Reproduce “ARMOR: Robust and Efficient CNN-Based SAR ATR through Model-Hardware Co-Design” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sachini Wickramasinghe, Tian Ye, Cauligi Raghavendra, Viktor Prasanna
A model-hardware co-design framework, ARMOR, enables robust and efficient Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) by integrating adversarial training with hardware-guided model compression and FPGA acceleration. It delivers up to 68.1x lower inference latency and 169.7x better energy efficiency on FPGAs compared to CPU baselines, while preserving adversarial robustness.
Key insights
- Hardware-guided pruning, by incorporating analytical hardware performance models, achieves superior robustness-efficiency trade-offs compared to pruning based on model metrics alone.
- First-order Taylor saliency effectively preserves adversarial robustness during structured pruning for SAR ATR models.
- FPGA accelerators, with channel-aware Processing Element (PE) allocation, provide significant latency and energy efficiency gains for compressed, robust models over CPU/GPU platforms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.