PRISM-TS: A Personalized Backdoor Defense Method for Power Load Time-Series Classification

For targeted backdoor attacks against power-load time-series classification models, attackers can exploit covert temporal triggers to misclassify genuine high-load states as low-load states, leading to issues such as underestimation of load risks and failure of alarms. Since trigger effects may vary across different network modules, existing methods，such as pruning, fine-tuning, and unlearning，typically apply a one-size-fits-all approach, making it difficult to balance backdoor suppression with the accurate identification of normal loads. To the best of our knowledge, this paper is the first to systematically investigate backdoor defense in this scenario, focusing primarily on local time-domain triggers and verifying the method’s applicability under frequency-domain attack conditions. To this end, we propose PRISM-TS, a proxy-trigger-guided personalized backdoor defense method. This method utilizes independent proxy triggers to assess module backdoor sensitivity and jointly selects high-risk modules under clean accuracy constraints. Differentiated cleansing strategies are applied to different modules: high-risk modules are replaced with corresponding modules from the clean reference model, while the remaining modules undergo weighted averaging between the clean and backdoor models. In random time-domain attack experiments on the AEP dataset, PRISM-TS reduced the attack success rate to 2.53% while maintaining a clean accuracy of 95.72% and a high-load retention rate of 89.62%. Frequency-domain extension experiments further demonstrate that this method offers effective defense capabilities.