Large-scale neural network compensation associated with camouflaging in trait autism and its potential mental health costs

Abstract

Background

Social camouflaging refers to strategies to hide or compensate for social difficulties, often at significant mental health costs, and is particularly prevalent in autism. The large-scale neural network associated with this adaptation remains poorly understood. This study aimed to identify these neural network patterns and their link to potential mental health issues.

Methods

Using a dimensional approach, we recruited 110 healthy young adults who completed self-report questionnaires measuring autistic traits and camouflaging as well as depression and anxiety, and underwent resting-state fMRI scans. The interaction between camouflaging and autistic traits on brain network connectivity was examined using the 300-node Seitzman atlas, encompassing 13 functional networks.

Results

Among individuals with higher autistic traits, greater camouflaging was associated with increased connectivity between the Default Mode Network (DMN) and the Cingulo-Opercular Network (CON), as well as within the CON. Crucially, DMN-CON hyperconnectivity statistically mediated the relationship between camouflaging and potential mental health costs (i.e., depression and anxiety scores) but only in individuals with higher autistic traits. Limitations: Our study was limited by its predominantly non-clinical sample, the cross-sectional design, and the use of resting-state rather than task-based fMRI.

Conclusions

These findings reveal specific compensatory neural network patterns associated with camouflaging in those high in autistic traits, involving interoception, self-referential, and executive control systems, and provide a neurobiological explanation for its potential mental health burden, highlighting the need for societal changes that reduce the pressure for such adaptations.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13229-026-00710-7.

Introduction

Atypical social communication and interaction, as well as restricted, repetitive behaviors or interests, are core symptoms of autism in clinical contexts [1]. Autistic traits share these qualitative features and are continuously distributed across the general population [2, 3], and even non-clinical individuals high in such traits often face similar social challenges [4, 5]. To better adapt to the demands of everyday predominantly neurotypical social contexts, some autistic individuals or those high in autistic traits consciously or unconsciously adopt strategies, known as social camouflaging, to minimize the visibility of their differences and challenges [6, 7]. While camouflaging may facilitate short-term social acceptance, it is also linked to diagnostic challenges (e.g., delayed, missed, or incorrect diagnosis) [8] and adverse mental health outcomes (e.g., anxiety, depression, and suicide ideation) [9, 10].

Previous qualitative and quantitative studies have shown that social camouflaging is a multidimensional construct, comprising three core components: masking (hiding autistic traits), compensation (using cognitive strategies to overcome social difficulties), and assimilation (striving to blend in) [6, 11]. Based on this framework, the Camouflaging Autistic Traits Questionnaire (CAT-Q) [12] was developed and is now widely used to measure social camouflaging. Recent theoretical studies indicated that camouflaging is not specific to autistic individuals or those high in autistic traits, but reflects a more general human tendency for impression management [13, 14].

However, given that the core features of autism involve social communication difficulties, social camouflaging in autistic individuals still exhibits unique characteristics (e.g., motivations, neurocognition, and consequences; see details in [13]). It often occurs without full insight into neurotypical social norms and involves distinct high-cost psychological processes to compensate for the core symptom [13]. As a ‘compensatory cost’, autistic individuals frequently experience fatigue and feelings of inauthenticity from camouflaging compared to non-autistic individuals [6], which may contribute to worse mental health [10, 15–17]. Hence, camouflaging-related neural compensation may vary with autistic traits, with potentially distinct cognitive neural patterns and mental health consequences in individuals with higher versus lower autistic traits, and remains poorly understood.

Several empirical studies have demonstrated that the main social symptoms of autism can be partially alleviated through functional compensation, typically by utilizing alternative compensatory neural pathways [18–20]. As a form of social compensation, camouflaging in autism may also be associated with some specific compensatory cognitive and neural functions. At the cognitive level, camouflaging in autism has been linked to enhanced executive function [11, 21, 22]. At the neural level, camouflaging in autistic females has been associated with increased activity in the medial prefrontal cortex (mPFC), a core node of the Default Mode Network (DMN), during self-representation [23], as well as with the excitation-inhibition balance within this region during resting-state [24]. Functional and structural connectivity patterns related to camouflaging have also been observed within circuits implicated in reward, emotion, and memory retrieval [25]. These findings suggest that camouflaging in autistic individuals or those high in autistic traits is not supported by separate brain regions but by interconnected neural circuits, highlighting the need for further investigation from a large-scale network perspective.

To address this, the current study adopted a dimensional approach to autistic traits and camouflaging. It analyzed resting-state fMRI data from 110 healthy adults using the 300-node Seitzman atlas [26], which incorporates subcortical and cerebellar regions into 13 functional networks, to investigate the specific neural network correlates of camouflaging across higher versus lower levels of autistic traits. Based on previous neuroimaging findings, we primarily focused on the DMN and networks supporting executive function (e.g., fronto-parietal network, FPN; cingulo-opercular network, CON) [27, 28], as well as their dynamic switching hub - salience network (SN) [29–31]. We hypothesized that the camouflaging × autistic trait interaction, reflecting how autistic traits modulate the neural correlates of camouflaging, would be associated with inter- or intra-network connectivity within the DMN, FPN, CON, and SN. We further examined whether these camouflaging-related neural networks were linked to potential mental health costs using moderated mediation models. Specifically, we hypothesized that these camouflaging-related neural circuits might operate in a high-cost compensatory manner in the context of higher autistic traits, thereby mediating the association between camouflaging and adverse mental health outcomes.

Materials and methods

Participants

The final sample for the current study consisted of 110 right-handed healthy adults (76 females; age range = 17–28 years, M = 19.93, SD = 1.54). These participants were recruited from an initial cohort of 124 college students in China. Fourteen individuals were excluded from the analysis due to missing data on the Camouflaging of Autistic Traits Questionnaire (CAT-Q) or excessive head motion during the resting-state scan (> 3 mm translation or > 3° rotation at any TR). The final sample showed good motion quality, with a mean framewise displacement (FD) [32] of 0.102 mm (SD = 0.054). All participants were free of any self-reported history of diagnosed mental or neurological disorders. The study protocol was approved by the local ethics committee at the Institute of Brain and Psychological Sciences, Sichuan Normal University (Approval No. 2024LS028) and conducted in accordance with the latest revision of the Declaration of Helsinki. All participants provided written informed consent before the experiment (parental consent was obtained for one participant under the age of 18) and received financial compensation.

Questionnaires

The camouflaging autistic traits questionnaire (CAT-Q)

The CAT-Q is a 25-item self-report instrument on which participants rate their behaviors on a 7-point scale (from 1 = strongly disagree to 7 = strongly agree) [12]. Five items are reverse-scored, and the total score is calculated by summing all item scores, with higher scores indicating greater use of camouflaging. In the present study, we used the Chinese translation of the original 25-item CAT-Q to assess camouflaging, which demonstrated good internal consistency in the current sample (Cronbach’s α = 0.829). This Chinese version was later psychometrically evaluated and refined into a 23-item version [33].

The autism-spectrum quotient (AQ)

The AQ is a 50-item self-report instrument on which participants rate statements on a 4-point scale (from 1 = definitely agree to 4 = definitely disagree) [34]. In the present study, we used the Chinese version of the AQ [35], which has demonstrated good psychometric properties in both clinical and non-clinical Chinese populations. Consistent with the original version, a binary scoring system was used: each item is converted to 0 or 1 depending on whether the response reflects the presence of autistic traits. The total score is calculated by summing all item scores, with higher scores indicating greater levels of autistic traits. The scale showed acceptable internal consistency in the current sample (Cronbach’s α = 0.715).

The Spielberger trait anxiety inventory (TAI)

The TAI is a 20-item self-report instrument on which participants rate how they generally feel on a 4-point scale (from 1 = rarely to 4 = almost always) [36]. After reverse-scoring the ten relevant items, the total score is calculated by adding all item scores, where higher scores reflect higher levels of trait anxiety. In the present study, the Chinese version of the TAI subscale, which has demonstrated good psychometric properties in Chinese young adults [37], was used to assess trait anxiety, showing excellent internal consistency in the current sample (Cronbach’s α = 0.885).

The Beck depression inventory (BDI-Ⅱ)

The BDI-Ⅱ is a 21-item self-report instrument on which participants rate the severity of their symptoms on a 4-point scale (from 0 to 3) [38]. The total score is calculated by summing the ratings, with higher scores indicating greater levels of depression. In the present study, the Chinese version of the BDI-Ⅱ, which has demonstrated good psychometric properties in Chinese populations [39], was used to assess depressive symptoms, showing excellent internal consistency in the current sample (Cronbach’s α = 0.910).

MRI data acquisition and preprocessing

MRI data acquisition and preprocessing steps are reported as recommended by the Organization for Human Brain Mapping [40]. MRI data were collected on a 3T Prisma MR scanner (Siemens, Erlangen, Germany) with a 64-channel head-neck coil at the Centre for MRI Research, Sichuan Normal University. During the scan, participants were instructed to rest with their eyes open, fixating on a blank screen, and remain as still as possible. Foam padding was used to minimize head motion. Resting-state functional MRI (rs-fMRI) data were acquired using a T2*-weighted echo-planar imaging (EPI) sequence (repetition time = 2000 ms, echo time = 30 ms, flip angle = 90°, FOV = 224 mm × 224 mm, 62 axial slices, slice thickness = 2.0 mm, slice gap = 0.3 mm, voxel size = 2 × 2 × 2.3 mm, multi-band factor = 2). The resting-state scanning session lasted approximately 8 min and consisted of 242 volumes. A high-resolution T1-weighted anatomical image was also acquired to improve the normalization of the functional images (repetition time = 2530 ms, echo time = 2.98 ms, flip angle = 7°, matrix size = 224 × 256, 192 sagittal slices, slice thickness = 1.0 mm, voxel size = 0.5 × 0.5 × 1 mm).

Preprocessing of resting-state fMRI data was performed using the Data Processing Assistant for Resting-State fMRI (DPARSF v5.4_230110) [41], a toolbox based on SPM12 and DPABI [42]. The initial five volumes were discarded to allow for scanner signal stabilization. The remaining functional images underwent slice-timing (reference slice: 32, the middle slice) and realignment for head motion correction using a six-parameter rigid body algorithm. Structural T1 images were segmented and bias-corrected. Functional images were then co-registered to the individual’s T1 image before being normalized to the Montreal Neurological Institute (MNI) standard space and resampled to an isotropic 2 × 2 × 2 mm³ voxel size. Subsequently, the data were cleaned by regressing out several nuisance covariates [43, 44], including Friston 24 head motion parameters (the standard 6 motion parameters and their temporal derivatives, quadratic terms, and squares of derivatives) [45], as well as mean signals from white matter, cerebrospinal fluid, and the whole brain (global signal). The residual data were then linearly detrended, band-pass filtered (0.01–0.1 Hz), and spatially smoothed with a 6-mm FWHM Gaussian kernel.

Functional network analysis

Functional brain networks were constructed for each participant by first defining connectivity at the level of regions of interest (ROIs) and subsequently summarizing these into network-level metrics. To establish ROI-level connectivity, each participant’s brain was parcellated into 300 ROIs using the Seitzman atlas [26]. This atlas assigns each ROI to one of 13 functional networks (e.g., DMN, FPN, CON, SN), excluding the ‘unassigned’ category (see the Greene lab website https://greenelab.wustl.edu for full ROI list and MNI coordinates). After extracting the mean time series from each ROI, a 300 × 300 connectivity matrix was generated for each participant by computing Fisher’s Z-transformed Pearson’s correlations between all ROI pairs. To focus our analysis on synchronous neural activity and avoid the ambiguous interpretation of negative correlations, all negative correlation values were set to zero. To ensure the robustness of our findings, this process was repeated across multiple sparsity thresholds (10%, 30%, 50%, 70%, and 90%), retaining only the strongest positive connections at each level.

These detailed ROI-level matrices were then summarized to create the network-level metrics used in our primary analysis. Specifically, intra-network connectivity was calculated by averaging the correlation values among all unique node pairs within a given network, while inter-network connectivity was derived from the average correlations of all node pairs connecting two different networks. This procedure yielded a 13 × 13 summary matrix of mean functional connectivity for each participant at each sparsity threshold. The diagonal elements of this matrix represent intra-network connectivity, and the off-diagonal elements represent inter-network connectivity.

A General Linear Model (GLM) was conducted on the network-level connectivity matrices to test for the effects of camouflaging (CAT-Q), autistic traits (AQ), and their interaction term (AQ × CAT-Q) on each of the intra- and inter-network connections, while controlling for age and sex. The interaction term captures how individuals’ autistic traits (AQ) modulated the neural correlates of camouflaging (CAT-Q), beyond the main effects of either variable alone. To ensure the stability of the regression model, AQ and CAT-Q were z-standardized before computing the interaction term, and Variance Inflation Factor (VIF) diagnostics indicated no multicollinearity issues (all VIFs < 2). The resulting p-values for the 13 intra-network and 78 inter-network connections were independently corrected for multiple comparisons using the False Discovery Rate (FDR) procedure (pFDR < 0.05), which is recommended in neuroimaging studies [46, 47]. For robustness, we prioritized network connections that remained significant across all tested sparsity thresholds.

To visualize and localize the specific connections and flexible hubs driving significant network-level effects, we subsequently utilized the Network-Based Statistics (NBS) toolbox [48] as a complementary edge-level analysis. For each significant network-level finding, NBS was used to identify a connected component (i.e., a cluster of edges) whose connectivity strength was significantly associated with the interaction between CAT-Q and AQ (pFWE < 0.05, component-level correction based on 10,000 permutations). All statistical analyses were performed in MATLAB R2018a (The MathWorks, Inc.).

Associations between functional networks and mental health outcomes

Preliminary analyses included descriptive statistics for all variables by sex. Sex differences and bivariate correlations were assessed using independent t-tests and Pearson’s r, respectively, with Bayes factors (BF₁₀) reported [49]. No adjustments for multiple comparisons were made since these analyses were purely descriptive and exploratory.

To identify candidate neural mediators for subsequent moderated mediation analyses, we first examined the partial correlations between the significant neural networks from previous analyses and the mental health measures (depression and anxiety), with age and sex included as covariates. Given the exploratory nature of this step and the need to balance Type I and Type II error rates, FDR correction was applied for multiple comparisons. Neural pathways showing significant associations with mental health were subsequently entered into moderated mediation models, whose robustness was evaluated using bias-corrected bootstrap confidence intervals.

Moderated mediation analyses were used, theoretically based on the ‘compensatory cost’ of autistic camouflaging [6, 13], to investigate whether the identified neural pathways mediated the relationship between social camouflaging and mental health using the PROCESS macro [50] in SPSS (Model 7). In these models, we tested whether the indirect effect of camouflaging (CAT-Q; independent variable, X) on mental health outcomes (anxiety or depression; dependent variable, Y) via the significant neural pathway (mediator, M) was moderated by autistic traits (AQ; moderator, W). The analysis was based on 5,000 bootstrap samples, and 95% confidence intervals (CIs) were generated to test the significance of the conditional indirect effects and the index of moderated mediation. To further identify the specific range of autistic traits (AQ scores) for which the relationship between camouflaging and the neural mediator was statistically significant, the Johnson-Neyman (J-N) technique was employed. The resulting J-N plot was generated using the Jnplots package [51] in the R statistical environment (version 4.4.1).

Results

Camouflaging is associated with autistic traits and poorer mental health

Descriptive statistics, sex differences, and bivariate correlations for all study variables are presented in Table 1. While the sample did not include clinical participants, 17 of 110 participants (≈ 15.5%) scored ≥ 30 on the AQ, 9 participants (≈ 8.2%) scored ≥ 60 on the TAI, and 2 participants (≈ 1.8%) scored ≥ 30 on the BDI-Ⅱ [52, 53]. Note that the thresholds mentioned here are provided for research reference purposes only, not for clinical diagnosis or screening.

Table 1.

Descriptive statistics, sex differences, and bivariate correlations for all variables

Variables	Descriptive Statistics M ± SD (Range)				Sex Differences t (BF10)		Pearson Correlations r (BF10)
	Total n = 110	Female n = 76	Male n = 34				1	2	3	4
1 Age	19.93 ± 1.54 (17–28)	19.62 ± 1.28 (17–24)	20.62 ± 1.84 (18–28)		3.29** (22.84)
2 CAT-Q	108.55 ± 15.31 (64–141)	111.30 ± 14.55 (72–141)	102.38 ± 15.37 (64–130)		2.92** (8.70)		-0.15 (0.38)
3 AQ	24.17 ± 6.02 (10–37)	25.04 ± 5.88 (11–37)	22.24 ± 5.95 (10–33)		2.30* (2.19)		-0.17 (0.57)	0.29*** (10.93)
4 TAI	46.45 ± 9.06 (22–66)	47.91 ± 9.59 (22–66)	43.21 ± 6.79 (31–62)		2.93** (3.92)		-0.12 (0.25)	0.31*** (24.90)	0.58*** (> 100)
5 BDI-Ⅱ	11.50 ± 8.80 (0–34)	12.47 ± 8.61 (0–34)	9.32 ± 8.96 (0–30)		1.75 (0.83)		-0.14 (0.34)	0.25*** (3.29)	0.38*** (> 100)	0.68*** (> 100)

Note. CAT-Q = Camouflaging Autistic Traits Questionnaire; AQ = Autism-Spectrum Quotient; TAI = Trait Anxiety Inventory; BDI-Ⅱ = Beck Depression Inventory-Ⅱ; M =Mean; SD = Standard Deviation; BF₁₀ = Bayes factor for the alternative hypothesis (H₁) relative to the null (H₀). *p < 0.05, **p < 0.01, ***p < 0.001 (two-tailed)

Sex differences were observed in several variables. Specifically, females scored significantly higher than males on the CAT-Q (t(108) = 2.92, p = 0.004, Cohen’s d = 0.60, BF₁₀ = 8.70) and TAI (t(87.48) = 2.93, p = 0.004, Cohen’s d = 0.57, BF₁₀ = 3.92), with moderate to strong Bayesian support. Although the sex difference in the AQ reached significance at the frequentist level (t(108) = 2.30, p = 0.023, Cohen’s d = 0.47), the corresponding Bayesian evidence was only anecdotal (BF₁₀ = 2.19), suggesting insufficient support for a robust effect. No significant sex difference was found for the BDI-Ⅱ (t(108) = 1.75, p = 0.083, Cohen’s d = 0.36, BF₁₀ = 0.83). In addition, males in the sample were significantly older than females (t(108) = 3.29, p = 0.001, Cohen’s d = 0.68, BF₁₀ = 22.84). Thus, sex and age were included as covariates in subsequent analyses to control for their potential confounding effects.

Pearson correlation analysis revealed a significant positive association between camouflaging (CAT-Q) and autistic traits (AQ; r = 0.29, p = 0.002, BF₁₀ = 10.93), indicating that individuals with higher autistic traits reported more camouflaging behaviors. Camouflaging was also significantly associated with poorer mental health, showing positive correlations with both depression (BDI-Ⅱ; r = 0.25, p = 0.010, BF₁₀ = 3.29) and trait anxiety (TAI; r = 0.31, p = 0.001, BF₁₀ = 24.90). Furthermore, autistic traits (AQ) were also significantly correlated with depression (BDI-Ⅱ; r = 0.38, p < 0.001, BF₁₀ > 100) and anxiety levels (TAI; r = 0.584, p < 0.001, BF₁₀ > 100). These behavioral results indicated the core associations between social camouflaging, autistic traits, and mental health in the current sample with moderate to decisive Bayesian support.

Interaction effect of camouflaging and autistic traits on network connectivity

To construct the functional connectivity matrices, all negative connections were excluded, and multiple sparsity thresholds (10%, 30%, 50%, 70%, and 90%) were applied to the positive connectivity matrices to ensure the robustness of our results. The proportion of zero-value entries in each participant’s network-level connectivity matrix at each threshold was evaluated. A network-level connection (e.g., DMN-CON inter-network FC) was defined as zero if all constituent ROI-to-ROI correlations were set to zero by the thresholding procedure. As shown in Supplementary Table S1, the 10% sparsity threshold resulted in an excessive proportion of zero-value entries, indicating a significant loss of network information and potential fragmentation of the network topology. Consequently, this threshold was excluded from further statistical analyses. The remaining sparsity thresholds (30%, 50%, 70%, and 90%) yielded an acceptable proportion of zero-value entries and were retained for all subsequent analyses. To maximize the signal-to-noise ratio and clearly delineate the most critical neural circuits, we primarily report the results from the 30% sparsity threshold in the main text, and the rest of the sparsity results are presented in the supplementary material.

A General Linear Model (GLM) was used to examine the effects of social camouflaging (CAT-Q), autistic traits (AQ), and their interaction on the 13 intra-network and 78 inter-network functional connections, while controlling for age and sex. The analysis revealed a significant and robust interaction between CAT-Q and AQ scores (see Fig. 1A & B). This interaction indicates that the relationship between camouflaging and brain network connectivity varies depending on autistic traits. Specifically, across all tested sparsity thresholds (30%, 50%, 70%, and 90%; see Supplementary Table S2), the interaction term significantly and positively predicted functional connectivity between the Default Mode Network (DMN) and the Cingulo-Opercular Network (CON) (30% sparsity: B = 0.010, t = 3.401, pFDR = 0.037; 50% sparsity: B = 0.008, t = 3.341, pFDR = 0.023; 70% sparsity: B = 0.007, t = 3.386, pFDR = 0.026; 90% sparsity: B = 0.005, t = 3.336, pFDR = 0.046), as well as intra-network connectivity within the CON (30% sparsity: B = 0.014, t = 3.573, pFDR = 0.007; 50% sparsity: B = 0.011, t = 3.320, pFDR = 0.008; 70% sparsity: B = 0.010, t = 3.348, pFDR = 0.015; 90% sparsity: B = 0.009, t = 2.920, pFDR = 0.028). Simple slopes analyses revealed that among individuals with higher autistic traits, greater social camouflaging was associated with stronger DMN-CON inter-network and CON intra-network functional connectivity (see Fig. 2C for details). These two brain network connectivity patterns were robustly associated with the CAT-Q × AQ interaction when fieldmap-corrected [54] neuroimaging data were used (see Supplementary Table S3).

Fig. 1.

Neural correlates of the camouflaging × autistic traits interaction. (A) Functional connections showing a significant interaction effect (pFDR < 0.05) across four sparsity thresholds, with robust effects for DMN-CON between-network and CON intra-network connectivity. (B) T-value matrices of the interaction effect at each sparsity threshold. (C-D) Connected components identified by Network-Based Statistics (p < 0.05, FWE-corrected), showing enhanced connectivity between the DMN (red nodes) and CON (blue nodes), and within the CON. Brain nodes are displayed in sagittal (left) and axial (right) views. Abbreviations: DMN = Default Mode Network; VIS = Visual Network; FPN = Fronto-Parietal Network; REN = Reward Network; DAN = Dorsal Attention Network; VAN = Ventral Attention Network; SN = Salience Network; CON = Cingulo-Opercular Network; SMN-D = Dorsal Somatomotor Network; SMN-L = Lateral Somatomotor Network; AUD = Auditory Network; PMN = Parieto-Medial Network; MTL = Medial Temporal Lobe Network

Fig. 2.

Moderated mediation models. (A-B) Indirect effects of social camouflaging (CAT-Q) on trait anxiety (TAI) and depression (BDI-Ⅱ) through DMN-CON functional connectivity, moderated by autistic traits (AQ). Unstandardized regression coefficients are shown for each path. (C) Johnson-Neyman (J-N) plot illustrating the conditional effect of social camouflaging on DMN-CON functional connectivity across the range of AQ scores. Orange and blue regions represent significant positive and negative associations, respectively, while the gray region indicates non-significance. Solid and dashed vertical lines mark the AQ values where the effect becomes or ceases to be statistically significant

In addition to this primary finding, other significant effects were observed at specific sparsity thresholds (see Supplementary Table S2). At the 50% and 70% thresholds, the interaction term also significantly predicted connectivity between the DMN and the Dorsal Attention Network (DAN), between the Ventral Attention Network (VAN) and both the Cingulo-opercular Network (CON) and Salience Network (SN), as well as intra-network connectivity within the Reward Network (REN). Regarding the main effect of CAT-Q, it was found to negatively predict intra-network connectivity within the REN at the 70% sparsity threshold (B = -0.011, t = -2.973, pFDR = 0.024) and robustly predicted lower intra-network connectivity within the Auditory Network (AUD) across all thresholds (Bs ranged from − 0.017 to -0.018; ts ranged from − 3.062 to -3.758; all pFDR < 0.036). No significant main effects of AQ scores were found.

Given the potential influence of sex on camouflaging, we further performed an exploratory GLM including the three-way interaction (Sex × CAT-Q × AQ) and all relevant lower-order terms, while controlling for age. This interaction was not significant after FDR correction across all retained sparsity thresholds (30%, 50%, 70%, and 90%; all pFDR > 0.423) or within any of the identified networks (DMN–CON inter-network connectivity and CON intra-network connectivity across all retained sparsity thresholds; all pFDR > 0.782; see Supplementary Table S4). These results suggest that sex did not significantly modulate the brain networks associated with the CAT-Q × AQ interaction.

To pinpoint the specific neuroanatomical connections driving the robust network-level findings, we employed Network-Based Statistics (NBS) and focused on the 30% sparsity threshold to isolate the strongest associations and clearly reveal the key neural circuits. The NBS analysis on DMN-CON between-network connections identified a significant connected component (28 edges connecting 28 nodes) positively associated with the CAT-Q × AQ interaction (pFWE < 0.05). This component was characterized by enhanced connectivity between core hubs of the DMN—including the pregenual Anterior Cingulate Cortex, Precuneus, Posterior Cingulate Cortex, and medial/inferior Orbitofrontal Cortex—and key regions of the CON, such as the Insula, Putamen, Thalamus, and Supplementary Motor Area (Fig. 1C). Similarly, the NBS analysis on CON intra-network connections revealed a significant connected component (144 edges and 30 nodes), also positively associated with the CAT-Q × AQ interaction (pFWE < 0.05). This component was extensive, encompassing all nodes within the Cingulo-Opercular Network (Fig. 1D), suggesting that a global increase in intrinsic CON coherence is a key neural correlate associated with this interaction.

The mediating role of compensatory neural pathways in mental health

For the two robustly significant neural markers (DMN-CON inter-network FC and CON intra-network FC at 30% sparsity), partial correlations with mental health measures (depression and anxiety) were examined while controlling for age and sex. To identify candidate neural mediators, FDR correction was applied. The results revealed that DMN-CON connectivity was significantly positively correlated with both trait anxiety (r = 0.22, p = 0.021, pFDR = 0.042) and depression (r = 0.23, p = 0.018, pFDR = 0.042). In contrast, the relationship between CON intra-network connectivity and mental health measures was non-significant (rs < 0.121; p > 0.213, pFDR > 0.284). Therefore, DMN-CON connectivity was carried forward as the mediator in our moderated mediation models. The analyses revealed significant overall indices of moderated mediation for both the anxiety model (Index = 0.069, 95% CI = [0.011, 0.145]; see Fig. 2A) and the depression model (Index = 0.072, 95% CI = [0.014, 0.155]; see Fig. 2B). These results suggested that the indirect effect of camouflaging on mental health via DMN-CON connectivity was significant and depended on the individual’s level of autistic traits. After accounting for this mediated pathway, the direct effects of camouflaging on both anxiety (B = 0.259, t = 2.801, p = 0.006) and depression (B = 0.209, t = 2.205, p = 0.030) remained significant. Detailed regression results are provided in Supplementary Table S5, with additional results at other sparsity thresholds reported in Supplementary Tables S6–S8.

To further specify the boundaries of this conditional effect, the Johnson-Neyman analysis revealed that for individuals with AQ scores above 29.72 (the top 15.46%; 17 of 110 participants, 3 males), camouflaging was significantly and positively associated with DMN-CON connectivity. Conversely, for individuals with AQ scores below 20.73 (the bottom 28.12%; 31 of 110 participants, 15 males), camouflaging was significantly and negatively associated with DMN-CON connectivity (see Fig. 2C). A chi-square test indicated a significant difference in sex distribution between the high- and low-AQ groups (χ² = 4.43, p = 0.035). Exploratory sex-stratified moderated mediation models suggested that the observed effects were primarily evident in females, whereas corresponding effects in males did not reach statistical significance, likely due to limited statistical power given the smaller male subsample (n = 34; details are provided in Supplementary Table S9).

Discussion

This study aimed to explore the large-scale brain networks associated with social camouflaging and its potential mental health costs in individuals with higher autistic traits. The results revealed a specific neural pattern of social camouflaging in those high in autistic traits, characterized by enhanced functional connectivity between the Default Mode Network (DMN) and the Cingulo-Opercular Network (CON), as well as increased intra-network connectivity within the CON. Furthermore, DMN-CON hyperconnectivity mediated the relationship between camouflaging and adverse mental health outcomes (anxiety and depression) in individuals high in autistic traits. These findings provide neural evidence consistent with a high-cost compensatory process in the context of high autistic traits, establishing a direct neurobiological link between the neural compensatory effort involved in camouflaging and its mental health consequences.

Brain network compensation of social camouflage associated with autistic traits

Our primary and most robust finding was the significant interaction between social camouflaging and autistic traits on both DMN-CON and intra-CON connectivities. The DMN, a non-specific task-negative network, is implicated in self-referential processing, social cognition, episodic and autobiographical memory, and mind wandering [55–58]. In contrast, the CON, recently renamed the action-mode network (AMN) [59], is a non-specific task-positive network that supports alertness, goal maintenance, sustained extrinsic attention, motor control, and feedback processing (e.g., pain, physiological states, cognitive errors), thereby enabling goal-directed behavior [59–62]. Typically, these two networks are segregated or anti-correlated, facilitating flexible switching between internal mentation and external tasks [57, 59, 63]. In this context, camouflaging-related intra-CON hyperconnectivity in individuals with higher autistic traits might reflect reduced efficiency within executive networks, necessitating sustained recruitment of interoceptive monitoring and executive control to continuously evaluate and regulate their social behaviors and maintain social scripts. Moreover, stronger camouflaging-related coupling between the DMN and CON in individuals high in autistic traits suggested reduced functional segregation alongside increased co-engagement of typically segregated systems. This indicated heightened goal-directed self-referential processing (i.e., actively monitoring and regulating self-presentation), which requires integration of internal self-referential processing with external social cues to guide adaptive social behavior.

Furthermore, the NBS analysis provided fine-grained insight into this pattern by identifying atypical connections both within all CON nodes and between core social-cognitive hubs of the DMN (the medial prefrontal cortex and posterior cingulate cortex) [64] and core interoceptive hubs of the CON (the anterior insula and thalamus) [65]. Notably, the DMN-CON hubs identified in the current study overlapped with core connecting hubs of the interoceptive system reported in recent studies using both 3T and 7T MRI [66, 67]. Previous research proposed that interoception plays a crucial role in social cognition by supporting bodily self-awareness and the integration of interoceptive and exteroceptive information, which together form the foundation for self-other distinction and theory of mind [68, 69]. From a developmental perspective, social functioning does not derive from inborn modules but rather develops iteratively through social dyads via interoceptive prediction and inference [70–72]. The computational model of camouflaging [13] proposed that social camouflaging can similarly be conceptualized as an iterative predictive coding system, in which individuals adjust moment-to-moment mentalizing and action options based on feedback from prior social interactions, with prediction errors driving recursive updates that guide adaptive social behavior. Although previous studies have reported mixed findings regarding interoceptive differences in autism [73, 74], our observation of hyper DMN-CON coupling and intra-CON coupling suggested that, in individuals with higher autistic traits, camouflaging might rely on enhanced interoceptive prediction and inference to compensate for atypical social functioning associated with autistic traits. Our findings provided empirical neural evidence for the computational model of camouflaging [13], extending it by demonstrating that large-scale brain systems involved in interoceptive prediction and inference are linked to the cognitively and physiologically high-cost nature of these compensatory adaptations.

Our findings offered an alternative, more adaptive perspective on the atypical connectivities between DMN and task-positive networks frequently reported in autistic research [75–78]. Rather than considering them as markers of structural or functional impairment (e.g., failed network segregation or cognitive rigidity) [30, 79, 80], our results linked them to the strategic behavior of social camouflaging, framing them as evidence of neurally effortful, plastic compensation. While previous studies have reported similar connectivity patterns, they might have relied on ambiguous network definitions or did not specifically examine the CON. Because the CON is spatially adjacent to other task-positive networks [59], including the SN and FPN, prior findings may have partially reflected CON-related effects that were not explicitly isolated. For example, enhanced connectivity between DMN and Central Executive Network (CEN, also referred to as FPN) was associated with better metacognitive and social skills in autistic individuals [75], where metacognitive strengths might compensate for specific social skill difficulties [75, 81, 82]; autistic females, who are commonly considered to engage more extensively in camouflaging [83], exhibited stronger DMN–CEN connectivity than both neurotypical females and autistic males [84].

Importantly, our findings highlighted a specific coupling between DMN and CON rather than FPN/CEN. Whereas the FPN/CEN primarily supports flexible, phasic problem-solving [85–87], the CON supports the stable, sustained monitoring and maintenance of goal-directed behavior [59, 60, 88]. According to the computational model of camouflaging [13], autistic individuals tend to adopt heuristic, low-resolution strategies (e.g., maintaining a fixed smile, imitating actions, or using scripted lines), explaining why autistic camouflaging relies more on ‘action maintenance’ than ‘flexible reasoning’. This distinction is critical as it reveals the specific neurocognitive demands of autistic camouflaging. Our robust finding of enhanced intra-CON connectivity provided additional neural evidence consistent with this model. By explicitly isolating the CON, our study clarified a network whose compensatory role may have been previously underestimated, thereby offering a more precise mapping of neural processes involved in camouflaging.

The mental health costs of neural compensatory camouflaging

Another key finding from our moderated mediation analyses was that, in individuals with higher autistic traits, DMN–CON hyperconnectivity positively mediated the association between camouflaging and poorer mental health outcomes. This pattern suggested that although DMN-CON hyperconnectivity might reflect the neural effort of camouflaging as an adaptive strategy, the chronic engagement of such a high-cost process would be associated with increased psychological burden, offering a potential explanation for why camouflaging is so frequently associated with anxiety and depression in the autistic and high-autistic-trait population [10, 15].

Consistent with this interpretation, previous work had shown that camouflaging in autism is linked to poor mental health, involving heightened self-monitoring, cognitive exhaustion, and feelings of inauthenticity [89]. Our findings aligned with this explanation and offered further neural-level support. The enhanced DMN-CON coupling, involving continuous interoceptive inference, self-monitoring, and goal maintenance [59, 88], reflected high-cost cognitive strategies associated with autistic traits-related adaptive behavior and mediated the association between camouflaging and mental health costs. This aligned with a recently proposed framework suggesting that anxiety in autistic adolescents stems from a continuous and intense hyperawareness of bodily signals, leading to sensory overload [90]. In sum, it is this sustained and neurally expensive compensatory activity that might manifest as anxiety and depression.

Intriguingly, the Johnson-Neyman analysis revealed that the neuro-cognitive profile of camouflaging differs at lower levels of autistic traits. In these individuals, greater camouflaging was associated with an opposite neural pattern: DMN-CON hypoconnectivity. This clearer functional segregation between the DMN and CON might reflect a more efficient, automated form of impression management, where self-referential and social cognitive processing is effectively decoupled from the execution of goal-directed social scripts [13]. Such relatively low-cost camouflaging might allow individuals to meet social expectations more effortlessly, thereby serving as a protective factor against anxiety and depression. This stark contrast suggested that camouflaging may involve different neurocognitive processes and associated costs across individuals with higher versus lower autistic traits.

Importantly, camouflaging also exerted a positive direct effect on mental health symptoms, such that the total effect of camouflaging remained positive even in individuals with lower autistic traits. This pattern suggested that, in individuals without pronounced autistic trait-related social difficulties, camouflaging might contribute to poorer mental health through a pathway distinct from high-cost neural compensation, such as negative affective experiences arising from inhibiting the authentic self. Prior research had shown that strong associations between camouflaging and mental health difficulties were primarily observed in autistic subgroups characterized by higher levels of negative affect or autistic traits [17], whereas in the general population, camouflaging was more closely linked to emotion regulation processes [91].

In addition, we observed a robust negative association between social camouflaging and intra-network connectivity within the Auditory Network (AUD), independent of autistic traits. From a resource allocation perspective, this sensory hypoconnectivity might represent a neural trade-off whereby resources are reallocated from primary sensory systems to support high-order processing such as emotional regulation (in those with low autistic traits) or interoceptive inference (in those with high autistic traits). Taken together, these findings suggested that while camouflaging at lower levels of autistic traits may rely on more efficient neural dynamics, it may still incur emotional costs through psychosocial pathways that are not driven by overload in large-scale network integration or segregation.

Notably, a recent cross-sectional study suggested that camouflaging in autistic individuals may have some beneficial effects on psychological well-being, but these effects were small [92]. Therefore, autistic camouflaging is more aptly viewed as a double-edged sword. Although our cross-sectional design prevents us from making causal claims, our findings have proposed an interesting hypothesis by revealing the neural pathway that links camouflaging to its adverse potential mental health consequences.

Limitations

This study has several strengths, including its robust network construction across multiple sparsity thresholds and the use of a sophisticated moderated mediation model to link brain, behavior, and mental health. However, some limitations must be acknowledged.

First, the current sample consisted of individuals without a formal clinical diagnosis, yet 17 of them scored above the AQ cut-off of 30. Notably, this cut-off approximated the significant threshold identified in our Johnson-Neyman analysis (AQ = 29.72), where AQ moderated the mediation between camouflaging, brain network connectivity, and mental health outcomes. These 17 participants completed a follow-up assessment using the Social Responsiveness Scale (SRS) [93], which indicated mild to moderate clinical risk (raw scores 72–118, T scores > 60), though they declined a formal DSM-5 evaluation due to stigma and personal concerns. This may imply the potential clinical relevance of our findings, while caution remains warranted when generalizing to clinically diagnosed populations. The current findings describe variation across the autistic trait continuum in the general population; future studies in a clinical sample are needed for confirmation.

Second, the current sample exhibited an imbalanced sex ratio, with a high predominance of females (approximately 2:1). Although sex was explicitly included as a covariate in all main analyses, supporting the robustness of our primary results, exploratory analyses revealed no significant sex-specific interactions in the neural network data. However, given the smaller number of male participants (n = 34), the moderated mediation effects were likely driven primarily by females. This sex imbalance may therefore limit the generalizability of our findings to the broader male population. Future studies with more balanced sex ratios are needed to confirm and extend the current findings.

Third, the cross-sectional design, while effective for identifying associations, cannot establish causality. Although our mediation model was grounded in the ‘compensatory cost’ framework of autistic camouflaging [6, 13], it relies on statistical associations derived from cross-sectional and observational data. As such, our models tested statistically consistent indirect pathways rather than causal mechanisms, and alternative model structures (e.g., connectivity as an outcome or shared vulnerability models) cannot be ruled out [94]. In addition, despite controlling for age and sex and accounting for baseline social difficulties via autistic traits (AQ), unmeasured confounding variables may still influence the observed associations. Longitudinal and experimental studies are needed to clarify the temporal and causal relationships among camouflaging, brain connectivity, and mental health outcomes.

Finally, behavioral assessments relied exclusively on self-report questionnaires. Although these instruments are well-validated, they are subject to limitations such as individual differences in introspection and susceptibility to social desirability bias. In addition, despite an adequate sample size for individual-differences analyses, resting-state functional connectivity measures remain vulnerable to uncontrolled physiological noise (e.g., vascular reactivity, motion, and respiration) [95]. Future studies incorporating physiological monitoring and test–retest designs will be important for further validating the neural specificity of the observed effects. Moreover, while the resting-state fMRI data reveal intrinsic network architecture, task-based fMRI would be needed to determine whether the observed DMN–CON hyperconnectivity is specifically engaged during active social camouflaging.

Conclusion

This study provides large-scale network evidence for a specific neural pattern associated with social camouflaging in individuals with higher autistic traits: hyperconnectivities between the default mode and cingulo-opercular networks, as well as intra-cingulo-opercular network. It indicated camouflaging involves a high-cost monitoring and regulation of both self-presentation and internal physiological states. By demonstrating how this neural effort is linked to the well-documented mental health burden of camouflaging, our work offers a neurobiological framework for this phenomenon. These findings may deepen our understanding of the immense effort expended by autistic individuals to navigate a neurotypical world and underscore the urgent need to drive societal change to reduce the pressure for such costly adaptations, thereby promoting the well-being of all individuals.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CAT-Q

Camouflaging Autistic Traits Questionnaire

AQ

Autism-Spectrum Quotient

TAI

Trait Anxiety Inventory

BDI-Ⅱ

Beck Depression Inventory-Ⅱ

DMN

Default Mode Network

VIS

Visual Network

FPN

Fronto-Parietal Network

REN

Reward Network

DAN

Dorsal Attention Network

VAN

Ventral Attention Network

SN

Salience Network

CON

Cingulo-Opercular Network

SMN-D

Dorsal Somatomotor Network

SMN-L

Lateral Somatomotor Network

AUD

Auditory Network

PMN

Parieto-Medial Network

MTL

Medial Temporal Lobe Network

Author contributions

Han Guo: Conceptualization, Methodology, Formal Analysis, Visualization, Writing – Original Draft Preparation, Writing – Review & Editing. Xiaobing Chen: Investigation, Data curation, Writing – Review & Editing. Aihua Zhou: Investigation, Writing – Review & Editing . Juan Kou: Conceptualization, Validation, Writing – Review & Editing. Yi Lei: Investigation, Resources, Writing – Review & Editing. Keith M. Kendrick: Conceptualization, Validation, Writing – Review & Editing. Lei Xu: Conceptualization, Methodology, Supervision, Funding Acquisition, Project Administration, Writing – Original Draft Preparation, Writing – Review & Editing.

Funding

This work was supported by the National Natural Science Foundation of China (grant numbers NSFC32500987, NSFC32100893) and the Sichuan Province Key Research and Development Project (grant number 2023YFWZ0003).

Data availability

The anonymized dataset is available on reasonable request from the corresponding author.

Declarations

Ethical approval and informed consent statements

The study protocol was approved by the local ethics committee at the Institute of Brain and Psychological Science, Sichuan Normal University (Approval No. 2024LS028) and was conducted in accordance with the latest revision of the Declaration of Helsinki. All participants provided written informed consent before the experiment (parental consent was obtained for one participant under the age of 18) and received compensation.

Competing interests

The authors declare no competing interests.