Abstract
This paper presents a fully decentralized federated adversarial vision transformer (ViT) framework for secure, privacy-preserving, and robust image-based malware classification. Unlike conventional federated learning that relies on centralized aggregation and remains vulnerable to privacy breaches and adversarial attacks, the proposed system employs blockchain-based decentralized aggregation integrated with secure multi-party computation. Encrypted local model updates are securely aggregated without a central server, while the blockchain ledger ensures transparency, tamper resistance, and trust. To further enhance security, a zero-knowledge proof-based mechanism validates masked model updates, enabling verifiable aggregation without exposing raw parameters. Clients reconstruct the global model through decentralized consensus, preventing direct access to others’ updates. Adversarial robustness is improved via client-side adversarial ViT training, incorporating projected gradient descent-generated malware images with clean samples, thereby reducing false classifications. Computational efficiency is achieved by leveraging pre-trained ViT variants for resource-constrained environments. Extensive experiments on Malimg, Microsoft BIG 2015, and Malevis datasets demonstrate superior performance, achieving accuracies of 98.30%, 98.93%, and 95.72%, respectively. Compared to centralized and federated adversarial ViTs, as well as state-of-the-art methods (FASe-ViT, FASNet, DM-Mal, Fed-Mal), the proposed framework consistently achieves higher accuracy, precision, recall, and F1-scores, while ensuring privacy, resilience, and decentralized trust.
