Lightweight Security-by-Design for TinyML Models in Constrained Embedded Environments

. 2026 Sep 9;26(18):5727. doi: 10.3390/s26185727

Algorithm 1 Security-Aware Compression Search

1:
Input: full-precision model $M_{0}$ , training set D, adversarial budget $ϵ$ , weights $α, β, γ$ , rounds R
2:
Output: compressed model $M^{*}$
3:
Initialize candidate pool $C \leftarrow {M_{0}}$
4:
for $r = 1$ to R do
5:
Generate pruned/quantized candidates $C_{r}$ from $C$
6:
for each candidate $M_{i} \in C_{r}$ do
7:
Train or fine-tune $M_{i}$ on D
8:
Compute $A c c_{clean} (M_{i})$ on clean validation split
9:
Compute $A c c_{robust} (M_{i})$ under FGSM at budget $ϵ$
10:
Compute $C R (M_{i})$ using Equation (2)
11:
Compute $S (M_{i})$ using Equation (4)
12:
end for
13:
Keep top decile of $C_{r}$ ranked by S, discard the rest
14:
$C \leftarrow$ surviving candidates
15:
end for
16:
$M^{*} \leftarrow \arg \max_{M_{i} \in C} S (M_{i})$
17:
return $M^{*}$