Backdoor Attacks
PulseAugur coverage of Backdoor Attacks — every cluster mentioning Backdoor Attacks across labs, papers, and developer communities, ranked by signal.
2 day(s) with sentiment data
-
New methods ensure backdoor attacks survive image compression
Researchers have developed new methods to ensure backdoor attacks remain effective even after images undergo lossy compression. The study highlights that standard compression techniques like JPEG can render malicious tr…
-
New defense tricks target backdoor attacks on face recognition systems
Researchers have developed a set of practical defense mechanisms to protect deep learning-based face recognition systems from backdoor attacks. These methods aim to safeguard systems where training data might be comprom…
-
Hugging Face study reveals backdoor attacks are rarely stealthy across all dimensions
A new study from Hugging Face investigates the stealthiness of backdoor attacks on deep neural networks, moving beyond traditional metrics like attack success rate and benign accuracy. The research proposes a "tri-space…
-
Researchers embed encrypted neural networks for AI backdoor attacks
New research details a method for constructing backdoor attacks that embed encrypted neural networks within host models. These hidden circuits can be decrypted and executed using a secret trigger, remaining difficult to…
-
New research tackles backdoor attacks in federated LLM fine-tuning
Two new research papers propose advanced methods to defend federated large language models (LLMs) against backdoor attacks. The first paper, FedGraph, introduces a graph-based aggregation framework that models relations…
-
New UniGuardian system detects multiple LLM attacks without prior knowledge
Researchers have introduced UniGuardian, a novel system designed to detect various attacks against large language models (LLMs) without prior knowledge of the attack type. This training-free detector identifies prompt i…
-
New NDDL framework defends text-to-image models against backdoor attacks
Researchers have developed a new defense framework called Normal Diffusion Dynamics Learning (NDDL) to combat backdoor attacks in text-to-image diffusion models. Unlike previous methods that rely on detecting specific a…
-
New methods tackle LLM backdoor attacks without retraining
Researchers have developed two new methods, NEEDLE and a null-space projection technique, to combat backdoor attacks in Large Language Models (LLMs). NEEDLE, a training-free approach, identifies and suppresses backdoor …
-
New framework tackles backdoor attacks in continual learning
Researchers have introduced a novel framework for continual learning that addresses the challenge of backdoor attacks in sequentially arriving tasks. This framework integrates sample purification, selective recovery, an…
-
New backdoor attack targets AI knowledge distillation process
Researchers have demonstrated a new method for backdooring image knowledge distillation, a process typically used to transfer capabilities from large AI models to smaller ones. The attack involves poisoning the distilla…
-
New black-box defense TRIM detects and removes AI backdoor triggers
Researchers have developed TRIM, a novel black-box defense system designed to detect and remove backdoor triggers in deep neural networks (DNNs) during inference. Unlike previous methods that require access to model int…
-
New PEPPER defense combats backdoor attacks in text-to-image models
Researchers have developed a new defense mechanism called PEPPER (PErcePtion-Guided Perturbation) to combat backdoor attacks in text-to-image diffusion models. These attacks can manipulate model outputs towards harmful …
-
New TRACE Method Detects Backdoors in Object Detection Models
Researchers have developed a new method called TRACE (TRAnsformation Consistency Evaluation) to detect poisoned samples in object detection models at test time. This technique addresses the unique challenges posed by ob…
-
New research details backdoor attacks and defenses for speech recognition models
Two new research papers explore the vulnerabilities of speech recognition models to backdoor attacks. The first paper introduces SpeechGuard, a system designed to detect and neutralize these attacks in real-time by iden…
-
New defense tackles backdoor attacks on self-supervised AI models
Researchers have introduced a novel defense mechanism called Platonic Representation Defense to combat backdoor attacks on self-supervised learning (SSL) models. This method operates in a black-box setting, meaning it d…
-
New BYORn Framework Defends LVLMs Against Backdoor Attacks
Researchers have developed a novel defense framework called BYORn (Bootstrap Your Own Responses) to protect Large Vision-Language Models (LVLMs) from backdoor attacks during supervised fine-tuning (SFT). This method lev…
-
Trigger color significantly impacts federated learning backdoor attack success
Researchers have demonstrated that the color of visual triggers significantly impacts the success rate of backdoor attacks in federated learning. By manipulating trigger colors on semantic objects like masks and sunglas…
-
New research details backdoor attacks on IoT/CPS continual learning systems
A new paper details a backdoor attack specifically designed for continual learning systems within Internet of Things (IoT) and Cyber-Physical Systems (CPS). The research highlights how continual adaptation, while benefi…
-
New RING attack exploits differential privacy in federated learning
Researchers have developed a new attack method called RING that exploits differential privacy (DP) in federated learning (FL) to conceal malicious updates. Contrary to prior assumptions, DP can mask the statistical char…
-
New methods tackle LLM backdoor attacks using shared mechanisms
Researchers have developed new methods to combat backdoor attacks in large language models (LLMs). One approach involves embedding a "dummy backdoor" to help remove unknown malicious triggers by fine-tuning the model on…