Detecting Poisoning Attacks in Quantized Federated Learning for IoT: A Trustworthy AI Approach

Split Learning (SL) enables collaborative model training across resource-constrained Internet of Things (IoT) devices while preserving data privacy, making it ideal for applications like smart homes and intrusion detection. However, SL systems are vulnerable to poisoning attacks, where malicious clients manipulate activations to compromise model integrity. This paper proposes a novel detection framework for poisoning attacks in quantized SL systems, leveraging pruned weights to enhance efficiency and an autoencoder-based method to identify malicious clients. Our approach uses intraclass-distance inflated loss to detect anomalies in quantized activations, tailored for IoT environments with device churn and data heterogeneity. Experimental results on the Bot-IoT dataset demonstrate high detection accuracy, with the autoencoder method achieving up to 80% true positive rates in random partitioning scenarios and maintaining model accuracy above 85% under attack. By integrating pruned weights and quantization, our framework reduces computational overhead by approximately 40% and communication costs by 80%, making it suitable for resource-constrained IoT devices. The approach aligns with Trustworthy AI principles by ensuring robustness, preserving privacy, and promoting fairness, offering a scalable solution for secure and ethical AI in IoT ecosystems.