Anomaly Detection in Autonomous Driving Systems for Identifying Adversarial Attacks : Evaluation of an Unsupervised Encoder-Decoder Convolutional Neural Network with Skip Connections

Sammanfattning

The rise of Autonomous Driving Systems (ADS) represents a significant advancement in automotive technology, promising enhanced safety, efficiency, and transformative impact on transportation. However, the increasing reliance on artificial intelligence and machine learning in ADS also introduces vulnerabilities to adversarial attacks. These attacks involve subtle manipulations of input data that can lead ADS to make incorrect decisions, posing substantial risks to safety and reliability. This thesis addresses the critical problem of detecting and mitigating such adversarial attacks in ADS. This research is particularly relevant given the complexity and significance of ensuring the security of ADS. Despite advancements, current methods often lack robust protection against sophisticated adversarial threats. The complexity of ADS environments and the subtlety of adversarial perturbations make this an ongoing challenge, underscoring the need for innovative solutions. To tackle this problem, this thesis implements and evaluates the Skip-GANomaly model, an unsupervised anomaly detection framework leveraging Generative Adversarial Networks (GANs). The model is trained on the BDD100k dataset and tested against various adversarial attack methods, including FGSM, BIM, PGD, DeepFool, and CW attacks. The methodology involves generating adversarial examples and assessing the model’s ability to detect these anomalies through reconstruction error analysis. The results demonstrate that Skip-GANomaly shows promise in identifying adversarial perturbations, particularly in detecting attacks such as FGSM. However, the model exhibits limitations when dealing with more complex attacks like DeepFool and CW, highlighting areas for future improvement. Key findings indicate that while Skip-GANomaly can enhance the robustness of ADS, further refinement and integration with additional defensive mechanisms are necessary to achieve comprehensive protection. This research has significant implications for the deployment of secure ADS. By enhancing anomaly detection capabilities, the findings can help develop more resilient autonomous systems, ultimately contributing to safer and more reliable autonomous transportation solutions. Future work could focus on optimizing model parameters, leveraging cloud-based resources for high-resolution training, and conducting real-world testing to validate the effectiveness of these solutions in practical scenarios. This research provides a foundation for further advancements in securing ADS against adversarial threats.