Published Paper Details:

This survey explores the evolving landscape of adversarial attacks on neural networks, with a focus on sponge attacks, which exploit the efficiency mechanisms of multi-exit networks (MENs) to maximize energy consumption and inference latency without degrading classification accuracy. Sponge attacks pose significant challenges to real-time applications like autonomous systems and healthcare diagnostics by negating the computational efficiency of MENs. Additionally, the survey examines the broader context of adversarial and backdoor attacks, such as data poisoning, which target model integrity during training. Recent advancements in defensive mechanisms, including certified defenses against poisoned data, defensive data augmentation, and anomaly detection, are analyzed for their effectiveness in addressing these threats. The role of real-time attack response systems (RTARS) and multi-exit architectures in mitigating latency-based attacks is also explored. Key findings highlight that while anomaly detection and pre-filtering techniques provide foundational security, gaps remain in handling hybrid and evolving adversarial threats like sponge poisoning. This survey concludes by identifying the necessity of integrating hybrid detection mechanisms--combining anomaly detection, real-time monitoring, and data augmentation--to enhance the robustness of neural networks against sponge attacks and other adversarial challenges, ensuring secure and efficient deployment in critical applications.