Tags → #backdoor attacks

HoneypotNet Develops Defense Against Model Extraction Attacks

HoneypotNet is a new defense mechanism designed to protect machine learning models from model extraction attacks through a multi-layered approach that includes backdoor detection and ownership verification methods.
Research Explores Backdoor Defense in Graph Neural Networks

Recent research explores data isolation techniques to enhance the security of Graph Neural Networks against backdoor attacks, highlighting methods for identifying and mitigating compromised samples during model training.
Post-Training Defense Method Enhances Malware Classifiers

Recent research has introduced a new post-training defense method, Post-Backdoor Purification, aimed at improving the security of machine learning models against backdoor attacks in malware classifiers.
Backdoor Attack Techniques in Outlier Detection Systems

Recent advancements in machine learning have introduced a tutorial on trigger design that can manipulate classifier performance, particularly in open-set scenarios.
Post-Training Backdoor Purification Method for Malware Classifiers

A recent tutorial discusses strategies to combat backdoor poisoning attacks on machine learning models used for malware classification, highlighting the Post-Training Backdoor Purification (PBP) method as a key countermeasure.
New Method Enhances Security of CLIP Models Against Backdoor Attacks

A new tutorial introduces the Perturb and Recover (PAR) method to enhance the security of Vision-Language models like CLIP against backdoor attacks, addressing critical vulnerabilities in AI systems.
FLARE Introduces Dataset Purification Against Backdoor Attacks

Recent advancements in artificial intelligence highlight the importance of securing deep neural networks against backdoor attacks, with a new tutorial introducing a defense mechanism called FLARE to help practitioners and researchers address these vulnerabilities.
Tutorial Initiative Launched for Personalized Learning Approaches

A new tutorial initiative has been launched to enhance educational resources through personalized and interactive learning approaches for students across various disciplines.
Large Language Models Used in Hardware Design Raise Security Concerns

Large language models are being increasingly used in hardware design for generating hardware description language code, but their integration raises significant security concerns related to data poisoning and backdoor attacks.
Study Reveals New Method for Backdoor Attacks on DNNs

A recent study has introduced Clean-Label Physical Backdoor Attacks (CLPBA), a method that employs physical triggers to manipulate machine learning models while maintaining the original labels of poisoned samples.