Delayed Backdoor Attacks: Exploiting the Temporal Dimension as a New Attack Surface in Pre-Trained Models
arXiv 2603.11949·high signal
Introduces Delayed Backdoor Attacks (DBA) that decouple malicious activation from trigger exposure—compromised models behave normally for N steps after trigger, then activate, defeating all detection methods that assume immediate backdoor expression. This temporal dimension was previously considered infeasible and invalidates snapshot-based auditing of pre-trained models. Any supply chain using third-party PTMs must now account for temporally-delayed behavioral compromise.