Machine learning models are not ciphers: extracting weights via power analysis is harder than expected
ML Models Are Not Ciphers: Power Analysis for Weight Extraction Is Harder Than You Think
Translation in progress — untranslated text is omitted.
The research questions the idea of treating machine learning models as ciphers and extracting all weights via power analysis, pointing out that full network weight recovery is harder than existing proofs of concept suggest. The authors assume a strong side-channel adversary with no noise that can directly observe the model's outputs, and still find two key differences between models and cipher implementations: models mostly use finite-precision floating-point arithmetic and low-diffusion linear operations, and data entropy decreases rather than increases as computation proceeds. These properties make deep weight recovery often require an unusually large number of traces. The research goal is a fully trained and working network, not merely recovering a small number of shallow weights or an untrained network.
- Recent work treats DNNs as cryptographic implementations and uses CPA to extract weights, but recovers only a small number of weights in shallow layers. The authors question this direct application and study what intrinsic challenges full extraction of a trained deep model faces.
- Build a simulation framework based on HW/HD leakage models, perform sequential CPA extraction weight by weight and layer by layer under noiseless and noisy conditions, analyze the difficulties caused by floating-point operations and decreasing data entropy, and propose strategies such as sign correction and adaptive receptive fields.
- Under noiseless conditions, AES can be recovered with fewer than 10 traces, an 8-layer CNN requires 5,000 traces, VGG-11 requires 20,000 traces, and ResNet-18 still fails with 200,000 traces; the approximate model can reach baseline accuracy with a small number of traces; at realistic noise levels, extraction fails at the first layer.
- The authors state that existing techniques are difficult to scale to larger models such as GPT-3, that under real noise the attack fails at the first layer, and that they believe entirely new attack techniques are needed; extraction difficulty for integer-arithmetic models is left for future work. The experiments are simulations; no real physical measurements were made.