Federated learning(FL)faces severe backdoor threats.Due to the inaccessibility of clean samples,the parameter server cannot clean them up in real time even if poisoning features are discovered.Meanwhile,existing backd...Federated learning(FL)faces severe backdoor threats.Due to the inaccessibility of clean samples,the parameter server cannot clean them up in real time even if poisoning features are discovered.Meanwhile,existing backdoor defense methods always require sacrificing model accuracy or increasing communication delay in exchange for better FL trustworthiness.To address these challenges,we propose a novel data-free and swift backdoor purification(DSBP)scheme based on multi-teacher adversarial distillation to effectively erase various backdoor variants in FL.DSBP treats the purification task as an adversarial game process between knowledge inheritance and backdoor inhibition by enforcing the student model to learn ensemble results of multiple teacher models on reconstructed clean samples,while being insensitive to synthetic poisoned samples.In DSBP,we utilize the self-similarity of poisoned features to optimize the trigger generator and accelerate the convergence of DSBP during the adversarial distillation process.We validate the effectiveness of DBSP by comparing it with four state-of-the-art defense methods against three backdoor variants on three datasets.The average attack success rate can be reduced from 96.6%to 2.3%with only 300 rounds.展开更多
The Heterogeneous Capacitated Vehicle Routing Problem(HCVRP),which involves efficiently routing vehicles with diverse capacities to fulfill various customer demands at minimal cost,poses an NP-hard challenge in combin...The Heterogeneous Capacitated Vehicle Routing Problem(HCVRP),which involves efficiently routing vehicles with diverse capacities to fulfill various customer demands at minimal cost,poses an NP-hard challenge in combinatorial optimization.Recently,reinforcement learning approaches such as 2D Array Pointer Networks(2D-Ptr)have demonstrated remarkable speed in decision-making by modeling multiple agents’concurrent choices as a sequence of consecutive actions.However,these learning-based models often struggle with generalization,meaning they cannot seamlessly adapt to new scenarios with varying numbers of vehicles or customers without retraining.Inspired by the potential of multi-teacher knowledge distillation to harness diverse knowledge from multiple sources and craft a comprehensive student model,we propose to enhance the generalization capability of 2D-Ptr through Multiple Teacher-forcing Knowledge Distillation(MTKD).We initially train 12 unique 2D-Ptr models under various settings to serve as teacher models.Subsequently,we randomly sample a teacher model and a batch of problem instances,focusing on those where the chosen teacher performed best.This teacher model then solves these instances,generating high-reward action sequences to guide knowledge transfer to the student model.We conduct rigorous evaluations across four distinct datasets,each comprising four HCVRP instances of varying scales.Our empirical findings underscore the proposed method superiority over existing learning-based methods in terms of both computational efficiency and solution quality.展开更多
基金supported by the National Natural Science Foundation of China under Grant Nos.62202303,U21B2019,and 62471301.
文摘Federated learning(FL)faces severe backdoor threats.Due to the inaccessibility of clean samples,the parameter server cannot clean them up in real time even if poisoning features are discovered.Meanwhile,existing backdoor defense methods always require sacrificing model accuracy or increasing communication delay in exchange for better FL trustworthiness.To address these challenges,we propose a novel data-free and swift backdoor purification(DSBP)scheme based on multi-teacher adversarial distillation to effectively erase various backdoor variants in FL.DSBP treats the purification task as an adversarial game process between knowledge inheritance and backdoor inhibition by enforcing the student model to learn ensemble results of multiple teacher models on reconstructed clean samples,while being insensitive to synthetic poisoned samples.In DSBP,we utilize the self-similarity of poisoned features to optimize the trigger generator and accelerate the convergence of DSBP during the adversarial distillation process.We validate the effectiveness of DBSP by comparing it with four state-of-the-art defense methods against three backdoor variants on three datasets.The average attack success rate can be reduced from 96.6%to 2.3%with only 300 rounds.
基金in part by the National Science Foundation of China under Grant No.62276238in part by the National Science Foundation for Distinguished Young Scholars of China under Grant No.62325602in part by the Natural Science Foundation of Henan,China under Grant No.232300421095.
文摘The Heterogeneous Capacitated Vehicle Routing Problem(HCVRP),which involves efficiently routing vehicles with diverse capacities to fulfill various customer demands at minimal cost,poses an NP-hard challenge in combinatorial optimization.Recently,reinforcement learning approaches such as 2D Array Pointer Networks(2D-Ptr)have demonstrated remarkable speed in decision-making by modeling multiple agents’concurrent choices as a sequence of consecutive actions.However,these learning-based models often struggle with generalization,meaning they cannot seamlessly adapt to new scenarios with varying numbers of vehicles or customers without retraining.Inspired by the potential of multi-teacher knowledge distillation to harness diverse knowledge from multiple sources and craft a comprehensive student model,we propose to enhance the generalization capability of 2D-Ptr through Multiple Teacher-forcing Knowledge Distillation(MTKD).We initially train 12 unique 2D-Ptr models under various settings to serve as teacher models.Subsequently,we randomly sample a teacher model and a batch of problem instances,focusing on those where the chosen teacher performed best.This teacher model then solves these instances,generating high-reward action sequences to guide knowledge transfer to the student model.We conduct rigorous evaluations across four distinct datasets,each comprising four HCVRP instances of varying scales.Our empirical findings underscore the proposed method superiority over existing learning-based methods in terms of both computational efficiency and solution quality.
