Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
This research paper presents a new method for training AI models when labels (correct answers) are ambiguous or uncertain. The method, called PLL-OVA (preferred-label partial-label learning in a one-vs-all view), improves on existing techniques by using ranked candidate labels instead of unordered sets, accounting for realistic annotation errors where the true label might be missing entirely, and introducing stability techniques called risk-correction functions (mathematical adjustments like ReLU that make training more robust).
This research paper presents a method for searching location-based services (apps that use your geographic position, like finding nearby restaurants) while protecting user privacy and ensuring the results are trustworthy. The approach combines spatio-temporal (location and time-based) keyword searching with verifiability (a way to prove the results are correct), allowing users to query location services without exposing their exact location or search patterns to the service provider.
This paper addresses backdoor attacks (where attackers secretly poison AI models to make them behave maliciously) in vertical federated learning (VFL, a setup where different organizations train an AI together on their own private data). The researchers propose a defense using a latent masked autoencoder (LMAE, a type of neural network that detects patterns and missing information) to identify when one participant is submitting suspicious, inconsistent data compared to honest participants, allowing the system to reject malicious contributions.
Fix: The paper proposes a novel defense mechanism using a latent masked autoencoder (LMAE) to assess the semantic consistency of embeddings (learned data representations) from different participants. The authors developed an algorithm based on the LMAE that identifies attackers and enables backdoor-resistant predictions. The defense was tested on multiple datasets and backdoor attack types and demonstrated effectiveness at identifying attackers while maintaining high prediction accuracy.
IEEE Xplore (Security & AI Journals)Diffusion models (AI systems that generate images and other content by gradually removing noise from random data) are vulnerable to backdoor attacks, where hidden triggers cause the model to produce harmful outputs. Researchers created PureDiffusion, a framework that can both defend against these attacks by detecting and inverting the hidden triggers, and amplify attacks by making existing backdoors more effective.
This paper presents sparse variational Student-t processes (SVTPs), a new machine learning method that improves how AI models handle data with outliers and extreme values. While Gaussian processes (GPs, statistical models that make predictions by learning patterns from data) are commonly used, they struggle with data that has heavy tails (rare extreme values far from the average), so the researchers developed SVTPs as a more robust alternative that also works efficiently on very large datasets.
The OWASP GenAI Security Project, an open-source community focused on AI security, announced expansion of its resources and frameworks with over 25,000 members contributing practical guidance and tools. The project is being highlighted at the RSA 2026 conference, indicating growing industry adoption of AI security best practices.
This survey examines methods for automatically finding bugs in software code by using machine learning and AI models, tracing the evolution from traditional machine learning techniques to modern large language models (LLMs, which are AI systems trained on vast amounts of text data). The research covers how these AI-based approaches learn patterns to pinpoint where faults occur in code, making debugging faster and more efficient than manual inspection.
This research addresses a weakness in active defense systems against deepfakes (AI-generated fake videos or images): these defenses often fail when attackers retrain their models on protected samples. The authors propose a Two-Stage Defense Framework (TSDF) that uses dual-function adversarial perturbations (carefully designed noise patterns that disrupt both the deepfake output and the attacker's retraining process) to make defenses more persistent by poisoning the data (corrupting the training information) that attackers would use to adapt their models.
Fix: The source describes the proposed defense framework (TSDF) as the solution but does not mention an existing patch, update, or mitigation for current systems. The paper presents the framework as a research contribution rather than a fix for deployed software. N/A -- no mitigation for existing systems discussed in source.
IEEE Xplore (Security & AI Journals)Decentralized Federated Learning (DFL, a way for multiple computers to train AI models together without a central server) is vulnerable to Byzantine attacks (when malicious participants send bad data to sabotage the learning process). The paper proposes FORCE, a new method that uses game theory concepts (mathematical models of strategy and fairness) to identify and exclude malicious clients by checking their model loss (how well their models perform) instead of checking gradients (the direction to improve), making DFL more resistant to these attacks.
This paper presents AdaptiveGDE, a new optimization algorithm designed to help train deep neural networks (DNNs, which are AI models with many layers) more effectively. The algorithm combines differential evolution (a method that mimics natural selection to find good solutions) with gradient descent (a technique that adjusts parameters to reduce errors), and uses an adaptive niching strategy (dynamically dividing the search into smaller groups) to balance exploring many possibilities early on with refining the best solutions later.
Android malware is a major security threat because the Android operating system's open app ecosystem allows unverified applications to be installed, making it easier for malicious software to spread and steal data, perform unauthorized financial transactions, or remotely control devices. Researchers are using machine learning (algorithms that learn patterns from data) to detect malware by analyzing features of Android application packages (APK files, the file format for Android apps), with recent research focusing on three main approaches: selecting the most important features to analyze, combining multiple detection models together, and handling datasets where malicious apps are much rarer than legitimate ones.
This academic paper is a systematic literature review (a comprehensive analysis of existing research) about physical unclonable functions, or PUFs, which are hardware-based security features that create unique, unchangeable identifiers for devices based on their physical properties. Published in July 2026, the review examines how PUFs are modeled and studied across different research papers. The paper does not describe a security problem or vulnerability, but rather surveys current knowledge about how these security devices work.
This is an academic survey paper published in ACM Computing Surveys that examines alignment of diffusion models (AI systems trained to generate images or other content by gradually removing noise from random data). The paper covers fundamental concepts, current challenges in making these models behave as intended, and directions for future research in this area.
This is a literature review article published in an academic journal that surveys how machine learning (algorithms that learn patterns from data to make predictions) is being applied to cybersecurity problems. The article covers research across the field but does not describe a specific security vulnerability or incident requiring a fix.
This is a survey article that reviews research on selective forgetting in machine learning, which is the ability to remove or reduce specific information from a trained AI model without completely retraining it from scratch. The article covers methods and applications of this technique across various AI systems and domains. The survey appears to be an academic overview of current knowledge in this area rather than describing a specific problem or vulnerability.
This academic review examines how bias (systematic unfairness in AI decision-making) occurs in AI systems and explores the human roles, solutions, and research methods used to identify and reduce it. The paper surveys existing approaches to addressing bias rather than proposing a single new solution.
This is an academic survey article published in ACM Computing Surveys that discusses a question bank designed to help assess risks in AI systems responsibly. The article appears to be a comprehensive review of how organizations can evaluate potential harms and safety concerns when developing or deploying AI, rather than describing a specific vulnerability or problem.
This academic paper is a systematic review published in ACM Computing Surveys that examines how trust works in artificial intelligence systems using established trust theory frameworks. The article analyzes trust in AI through theoretical lenses rather than addressing a specific technical vulnerability or problem.
This survey article reviews methods for detecting training data used to build large language models (LLMs, which are AI systems trained on massive amounts of text to generate human-like responses). The paper examines various techniques that researchers have developed to identify and extract information about what data was used to train these models, which is important for understanding model behavior and potential privacy concerns.
This research paper studies whether deepfake detection systems (AI tools that identify fake videos made to look real) are fair across different groups of people based on ethnicity, gender, and age. The study found that these detection systems often perform differently depending on the person's background, meaning they work better for some groups than others. The paper highlights that bias in deepfake detection is an important fairness problem that needs attention.