CNN-SHIELDED FEDERATED LEARNING: DUAL-PHASE ANOMALY DETECTION AND COMPRESSION AGAINST POISONING

Abstract

Poisoning attacks pose a significant threat in federated learning, where even a small fraction of malicious 
users (1–10%) can severely disrupt the model’s performance. Studies indicate that over 30% of real-world 
federated systems have encountered such attacks, leading to accuracy drops of up to 50%. This 
undermines the reliability of federated learning, especially in critical fields like healthcare and finance. 
Traditional centralized learning systems are also at risk due to their reliance on a single data storage point, 
increasing vulnerability to attacks and data breaches. Additionally, manual data handling is often 
inconsistent and error-prone. To mitigate these risks, a two-step defense approach is proposed that 
integrates data compression with a federated learning framework built on Convolutional Neural Networks 
(CNNs). The process begins with data preprocessing, which includes removing missing values, separating 
inputs and labels, applying standard scaling, and splitting the dataset into 90% training and 10% testing 
portions. Model compression is then used to reduce the size of user updates, saving bandwidth and 
concealing potential attack signatures. The proposed CNN-based federated learning model enhances 
accuracy from 87% (achieved with DNN) to 99%, offering robust defense against poisoning attacks and 
significantly improving overall performance.