Machine Learning Based Detection Of Data Poisoning Sponge Attacks In MultiExit Neural Networks

Abstract

Data Poisoning Based Sponge Attack on Multi-Exit Neural Architectures is a machine learning based security system developed to detect and prevent malicious attacks in neural network environments. Multi-exit neural architectures improve computational efficiency by generating early predictions through multiple exit points, but they are vulnerable to sponge attacks and poisoned data that increase energy consumption, inference time, memory usage, and CPU utilization. In this project, machine learning algorithms are used to analyze parameters such as packet size, duration, layer activation, memory usage, and inference time to identify different attack types. The system applies data preprocessing, feature extraction, classification, and prediction techniques to improve attack detection accuracy and enhance neural network security. The proposed model helps in maintaining reliable AI performance, reducing computational overhead, and improving the overall efficiency and robustness of intelligent systems against adversarial attacks.