Technology News and Headlines

NDSS 2025 – Compiled Models, Built-In Exploits

Researchers at NDSS 2025 unveiled a systematic study of bit‑flip attacks targeting deep‑learning executables compiled by modern DL compilers. Unlike prior work that focused on flipping weights within frameworks, the new approach exploits publicly known model structure embedded in the executable code, requiring no access to confidential weights. Their automated tool identifies vulnerable bits with 70% confidence, and on average only 1.4 flips are enough to reduce model accuracy to random guessing, even for quantized models. The findings expose a previously overlooked attack surface and call for security mechanisms in compilation toolchains.