Zero-Shot Deep Neural Network Model Extraction Across Heterogeneous Devices Using Side-Channel Translation and Foundation Models

Side-channel attacks have emerged as a seriousthreat to deep neural network (DNN) intellectual property (IP),especially on edge computing platforms. However, current attacksare limited to fixed hardware environments and do not generalisewell across heterogeneous devices. In this work, we proposea novel zero-shot model extraction framework that leveragescontrastive learning and large pre-trained models to translateside-channel traces across GPU architectures. Our method constructsa shared latent representation space that enables layersequence inference of DNNs from previously unseen hardwaretraces. We evaluate our approach using real-world GPU platforms\u2014NVIDIA RTX 3090, Jetson Nano, and AMD RadeonRX 580. Results demonstrate that our framework achieves upto 78.6% accuracy in cross-device DNN reconstruction, outperformingstate-of-the-art single-device attacks like DeepSniffer andTransformer-based baselines. To the best of our knowledge, thisis the first work that introduces cross-platform, zero-shot layerinference using side-channel signals, posing new challenges toDNN IP protection.