The field of network security and privacy is moving towards a greater emphasis on understanding and mitigating the risks of adversarial attacks and privacy leaks. Researchers are highlighting the importance of considering domain-specific constraints and model architecture when evaluating and designing robust models for security-critical applications. Additionally, there is a growing recognition of the limitations of traditional evaluation metrics for membership inference attacks and model inversion attacks, with a need for more robust and reliable frameworks. Notably, innovative methods such as the Loss-Based with Reference Model algorithm are being developed to improve the detection of memorization in generative and predictive models. Noteworthy papers include:

• Constrained Network Adversarial Attacks: Validity, Robustness, and Transferability, which reveals a critical flaw in existing adversarial attack methodologies.
• The DCR Delusion: Measuring the Privacy Risk of Synthetic Data, which shows that distance-based metrics fail to identify privacy leakage.
• Rogue Cell: Adversarial Attack and Defense in Untrusted O-RAN Setup Exploiting the Traffic Steering xApp, which introduces a detection framework to monitor malicious telemetry in O-RAN architectures.
• A new membership inference attack that spots memorization in generative and predictive models: Loss-Based with Reference Model algorithm, which proposes an innovative method to effectively extract and identify memorized training data.
• Uncovering the Limitations of Model Inversion Evaluation: Benchmarks and Connection to Type-I Adversarial Attacks, which presents an in-depth study of model inversion evaluation and reveals significant limitations in the widely used evaluation framework.