# Adversarial Machine Learning Attacks Against Video Anomaly Detection Systems

**Authors:** Furkan Mumcu, Keval Doshi, Yasin Yilmaz  
**arXiv:** [2204.03141](https://arxiv.org/abs/2204.03141)  
**Date:** 2022-04-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2204.03141v1/

## Key points

- Authors: Furkan Mumcu, Keval Doshi, Yasin Yilmaz from the University of South Florida
- Focus: Adversarial machine learning attacks against video anomaly detection systems
- Vulnerability: Surveillance cameras connected to anomaly detection servers via wireless networks
- Attack method: Wi-Fi deauthentication attack to generate adversarial data for video anomaly detection systems
- Experimentation: Applying effects like slowing down or reducing resolution on standard benchmark datasets using state-of-the-art models
- Findings: Attackers can trigger false alarms and conceal physical anomalies, undermining system reliability
- Tools mentioned: evillimiter for executing Wi-Fi deauthentication attacks effectively
- Impact: Compromising security measures and hindering accurate detection of genuine anomalies in video surveillance scenarios

## Summary

In their paper titled "Adversarial Machine Learning Attacks Against Video Anomaly Detection Systems," authors Furkan Mumcu, Keval Doshi, and Yasin Yilmaz from the University of South Florida address the critical issue of anomaly detection in videos within the realm of computer vision. The researchers delve into investigating an innovative adversarial machine learning attack specifically aimed at compromising video anomaly detection systems. They highlight the vulnerability of surveillance cameras connected to anomaly detection servers via wireless networks. The study demonstrates how a Wi-Fi deauthentication attack can be leveraged to generate adversarial data for video anomaly detection systems. By applying various effects induced by the attack on video quality, such as slowing down or reducing resolution, to standard benchmark datasets used for video anomaly detection, the researchers conduct experiments using state-of-the-art models. Their findings reveal that attackers can significantly undermine the reliability of these systems by triggering frequent false alarms and concealing physical anomalies. Furthermore, the paper sheds light on how adversaries can exploit tools like evillimiter to execute Wi-Fi deauthentication attacks effectively. The threat model outlined underscores the potential impact of such attacks on compromising security measures and preventing genuine anomalies from being accurately detected within video surveillance scenarios. This comprehensive exploration not only exposes critical vulnerabilities but also emphasizes the urgent need for robust defenses against adversarial machine learning attacks targeting video anomaly detection systems.