The field of deep learning security is moving towards developing more robust defenses against backdoor attacks. Recent research has focused on creating innovative solutions that can detect and mitigate these attacks, which involve manipulating a small subset of training data to cause misclassifications. One of the key directions in this area is the development of post-hoc defense methods that can scale across different types of triggers, including semantic triggers. Another important area of research is the analysis of the intrinsic vulnerabilities of deep models, which can be exploited to generate universal adversarial perturbations. Noteworthy papers in this area include: Prototype Guided Backdoor Defense, which proposes a robust post-hoc defense that scales across different trigger types. DeBackdoor, which presents a deductive framework for detecting backdoor attacks on deep models with limited data. Data-Free Universal Attack, which proposes a novel data-free method for generating universal adversarial perturbations by exploiting the intrinsic vulnerabilities of deep models.