VeCoSA-FI: A verifiable and collusion-resistant secure aggregation protocol in federated learning with fair incentive mechanism

As a key privacy-enhancing technology in distributed artificial intelligence, federated learning (FL) provides critical support for big data-driven applications while addressing privacy concerns. However, existing FL protocols face dual challenges of vulnerability to collusion attacks and lack of dynamic fair incentives, limiting their practical deployment in open ecosystems. To address these challenges, we propose a verifiable and collusion-resistant secure aggregation protocol with a fair incentive mechanism. Building upon the malicious secure aggregation model, our solution specifically counters collusion between server and initiator through homomorphic pseudorandom masking and linear homomorphic hashing. This ensures robust privacy preservation under comprehensive collusion threats while enabling efficient verifiable aggregation. Furthermore, we design a novel dynamic incentive mechanism that integrates user credibility with adaptive weight management, leveraging Pedersen commitments and Sigma protocols to guarantee weight correctness and privacy. Experimental results demonstrate that the proposed protocol maintains high computational efficiency and model accuracy, effectively resists free-riding behaviors among users, and introduces only acceptable additional communication overhead. Moreover, the protocol exhibits strong robustness to user offline events, effectively mitigating privacy leakage risks caused by collusion attacks while robustly preserving fairness in user collaboration. This work bridges advancements in information security, distributed computing, and AI, offering a scalable solution for trustworthy federated learning systems.