Beyond the Blind Spot: Deep Learning-Based Side-Channel Attack on MT-TMVP in Post-Quantum Cryptography

This deposit contains the power trace datasets and trained deep learning model weights accompanying the paper "Beyond the Blind Spot: Deep Learning-Based Side-Channel Attack on MT-TMVP in Post-Quantum Cryptography" (submitted to CHES 2026). Context The Modular Tiled Toeplitz Matrix-Vector Product (MT-TMVP) is a recently proposed FPGA-based polynomial multiplier for lattice-based post-quantum cryptography. This work presents the first side-channel security evaluation of this architecture. Power traces were captured from a Xilinx Artix-7 (XC7A100T) FPGA on a NewAE CW305 evaluation board using a ChipWhisperer-Lite (CW1173) at 7.37 MHz sampling rate (1 sample/clock, 10-bit ADC, 20 dB gain). The target implements NTRU polynomial multiplication with n=509, q=256, and ternary secret-key coefficients. Files File Size Contents sca_dataset_unprotected.zip 1.01 GB 59,049 power traces from the unprotected MT-TMVP implementation sca_dataset_protected.zip 1.22 GB 59,049 power traces from the protected MT-TMVP implementation (arithmetic masking + random delay + dummy rounds) models_unprotected.zip 10.27 MB 10 trained Deep Residual CNN checkpoints (.pth), one per target coefficient, for the unprotected design models_protected.zip 10.28 MB 10 trained Deep Residual CNN checkpoints (.pth), one per target coefficient, for the protected design Dataset Details Each dataset contains 59,049 traces corresponding to all 3^10 = 59,049 exhaustive ternary combinations of ten secret-key coefficients (f[0] through f[9]). Each trace is stored as a compressed NumPy archive (.npz) with the following fields: Key Shape Type Description wave (1, 20000) float64 Power trace (20,000 samples) labels (1, 10) int64 Ternary coefficient labels {-1, 0, 1} dut_io_ram_f_data (1, 509) int64 Secret polynomial f dut_io_ram_g_data (1, 509) int64 Known polynomial g dut_io_computed_data (1, 509) int64 Hardware-computed output The dataset is perfectly balanced: exactly 19,683 traces per class per coefficient. A fixed g polynomial (seed 42) is used across all traces. Model Details Each .pth file contains a PyTorch state dictionary for a Deep Residual CNN (~500K parameters) that classifies a single ternary coefficient from a raw 20,000-sample power trace. Architecture: Stem (Conv1d 1→64, k=15) → 4 Residual Blocks (64 channels, k=7, progressive MaxPool) → Head (Conv1d 64→128, k=5, AdaptiveAvgPool, Dropout 0.3, Linear 128→3). Training used AdamW with OneCycleLR, batch size 2048, label smoothing 0.1, and a 70/15/15 train/validation/test split (seed 42). Key Results Unprotected: 99.99% average per-coefficient accuracy, 99.91% all-correct rate (8,850/8,858 test traces) Protected: 94.66% average per-coefficient accuracy, 57.94% all-correct rate (42-percentage-point reduction) Replication The accompanying source code repository (hardware designs, capture scripts, training code, and analysis tools) is available at: https://anonymous.4open.science/r/Deep-Learning-Based-Side-Channel-Attack-on-MT-TMVP-E711 Keywords post-quantum cryptography, side-channel attack, deep learning, power analysis, FPGA, polynomial multiplication, Toeplitz matrix, NTRU, MT-TMVP