Autoencoder-enabled model portability for side-channel analysis
AI-generated audio on a research paper titled "Autoencoder-enabled model portability for reducing hyperparameter tuning efforts in side-channel analysis".
JCEN publication: https://link.springer.com/article/10.1007/...
eprint: https://eprint.iacr.org/2023/019
Keywords: Side-channel Analysis, Autoencoders, Preprocessing, Hyperparameter Tuning, Portability, Transfer Learning
Summary:
One of the main challenges in deep learning-based side-channel analysis (DLSCA) is hyperparameter tuning. For each side-channel dataset, hyperparameter tuning is done from scratch because side-channel measurements from different targets have different leakage. This means the same hyperparameters may not be equally efficient for different targets.
The paper explores using autoencoders for dimensionality reduction to see if encoded datasets from different targets enable the portability of profiling models and architectures. If successful, portability would mean:
Less effort is needed for hyperparameter tuning, as profiling model tuning would be eliminated for the new dataset.
However, tuning autoencoders remains, but it is more straightforward than tuning profiling models (see discussion in the paper).
The authors used three datasets:
DPAcontest v4.2 (DPAv4.2), which is an 8-bit Atmel ATMega-163 smart card protected with the Rotate Shift countermeasure.
ASCAD with a fixed key (ASCADf), which contains measurements from masked AES on the 8-bit ATMega8515 MCU target.
ASCAD with a random key (ASCADr), which has the same target and setup as ASCADf but with a variable key.
Experimental Setup
The authors performed a multi-step analysis:
Searching for the best latent space size: They conducted a random search on two datasets, DPAv4.2 and ASCADr, using four different types of autoencoders to find the best latent space size. Latent sizes of 200 and 400 were selected.
Selecting the best autoencoders: The best autoencoder was selected for each dataset based on the lowest mean squared error (MSE) for each latent space size and autoencoder type. They found that the best autoencoder types were Convolutional Neural Network (CNN) and Multi-Layer Perceptron (MLP), with a strictly decreasing number of neurons in subsequent layers, both using a latent size of 400.
Comparing the performance of profiling models trained with original and encoded traces: This step was performed to verify if the encoded datasets retained the relevant leakage information. The results showed that using original traces was not statistically significantly better than using encoded data, indicating that the encoded data preserved the necessary features for a profiling attack.
Investigating the portability of best profiling model hyperparameters: The authors tested the portability of the best profiling models trained with encoded data to other encoded datasets and the portability of the best profiling models trained with original data to other original and encoded datasets.
Investigating transfer learning of the best profiling model trained with original data to other encoded datasets: This was done to test the benefits of autoencoders in the context of transfer learning.
Results
The results showed that reusing profiling models trained on encoded data is possible. This reduces hyperparameter tuning efforts for new encoded datasets, suggesting that a universal profiling architecture might be possible for autoencoder-encoded data.
Furthermore, the authors found that encoded data was beneficial in cases where using original data for training resulted in unsuccessful attacks.
The authors conclude that using autoencoders to reduce the number of features in datasets can lead to better performance with lower tuning efforts compared to finding a new profiling model. They also found that transfer learning, enabled by autoencoders, worked quite well, eliminating the need to tune profiling models for new datasets and highly reduced training time.
Future Work
The authors suggest that future work should investigate CNN autoencoders further and provide a more in-depth analysis of the latent space representation of SCA traces using autoencoders. They propose comparing autoencoders with classical feature processing tools, such as principal component analysis (PCA). Finally, they suggest exploring the possibility of performing classical SCA on autoencoder-encoded data instead of using deep learning-based SCA.