The recent advancements in machine learning privacy and security have seen significant innovations, particularly in the areas of membership inference attacks, neuromorphic architectures, and backdoor detection. Researchers are increasingly focusing on developing methods that not only enhance privacy but also reduce computational overhead. Artifact-based privacy risk evaluation methods are emerging as a cost-effective alternative to traditional shadow model-based approaches, demonstrating high precision with minimal computational requirements. Neuromorphic architectures, such as Spiking Neural Networks (SNNs), are being explored for their inherent privacy-preserving properties, showing promise in mitigating data leakage risks compared to traditional Artificial Neural Networks (ANNs). Additionally, novel training approaches for SNNs, such as randomized forward mode gradients, are being developed to address the limitations of back-propagation, offering potential for more efficient and biologically plausible learning. In the realm of backdoor detection, scalable and efficient methods like Propagation Perturbation (ProP) are being proposed, which can identify malicious models without prior knowledge of triggers or malicious samples. These developments collectively indicate a shift towards more efficient, privacy-aware, and robust machine learning models.

Noteworthy papers include one that introduces a novel artifact-based approach for identifying at-risk samples with minimal computational overhead, and another that explores the privacy-preserving properties of SNNs, demonstrating their superior resilience against membership inference attacks compared to ANNs.