NDSS 2025 – BitShield: Defending Against Bit-Flip Attacks On DNN Executables

This piece highlights the NDSS 2025 paper "BitShield," a defense framework against bit-flip attacks (such as Rowhammer) targeting deep neural network executables. BitShield models attacks as semantic corruption and fuses code checksum routines into DNN semantics to detect manipulation of both data and executable sections, integrates with the TVM compiler, and claims broad applicability, proactive detection, and strong effectiveness (97.51% average mitigation) with low performance overhead (~2.47%).