CSMC: A Secure and Efficient Visualized Malware Classification Method Inspired by Compressed Sensing

View full-text article in PMC

. 2024 Jun 30;24(13):4253. doi: 10.3390/s24134253

Algorithm 2 The Proposed Classification Network Algorithm

Input:
Multi-dimensional tensor representations of malware samples.
1:
Initialize: Set up neural network parameters, layer configurations, initialize weights and biases.
Output:
Classification of malware into predefined families.
2:
for $b = 1$ to B epochs do
3:
${TensorList}_{b} \leftarrow SaveMultiDimVectors (^{'} . p t^{'})$ {Vectors from malicious software visualization}
4:
${TensorList}_{b} = {{Tensor}_{i} ∣ i = 1, \dots, n}$
5:
${Tensor}_{i} \in R^{b a t c h_s i z e \times c h a n n e l s \times h e i g h t \times w i d t h}$
6:
$TrainData \leftarrow convert_to_numpy ({TensorList}_{b} [1 : 0.9 n])$
7:
$AdjustTrainData \leftarrow reshape_for_network (TrainData)$
8:
$ValidationData \leftarrow convert_to_numpy ({TensorList}_{b} [0.9 n + 1 : n])$
9:
$ModelEvaluation (AdjustTrainData, ValidationData)$
10:
Transform s using compressed sensing to $μ \times s \times t$
11:
Network training over 30 layers:
12:
${Layer}_{0} : Input (μ \times s \times t \times c)$
13:
${Layer}_{1, 2} : Conv 2 D (3 \times 3, stride = 1, padding = ’ same ’)$
14:
${Layer}_{3} : MaxPool 2 D (3 \times 3, stride = 2, padding = ’ same ’)$
15:
${Layer}_{4 - 24} : ResidualShrinkageBlock (μ, s, t)$
16:
${Layer}_{25} : BatchNorm (decay = 0.9)$
17:
${Layer}_{26} : ReLU ()$
18:
${Layer}_{27} : GlobalAvgPool ()$
19:
${Layer}_{28} : FullyConnected (softmax)$
20:
${Layer}_{29} : Momentum (lr = 0.1, decay = 0.1, step = 20,000)$
21:
${Layer}_{30} : Regression (cross - entropy)$
22:
end for
23:
return Malware family $χ$