Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Bi3D++ is a machine learning method that helps 3D object detection systems (used in autonomous vehicles to identify objects like cars and pedestrians) work better when switching between different environments or sensors. Instead of requiring all new data to be labeled (annotated), it intelligently selects only the most important 1% of new data to label, then uses smart strategies to transfer knowledge from old labeled data to improve detection in the new environment.
This paper presents AFM-CIR, a framework designed to help AI models predict equipment failure times (RUL prediction) even when trained on data from one environment but tested on different environments (domain generalization). The approach combines Adaptive Fourier Mixing (a technique that blends training data in frequency space while preserving causal relationships) with Causality-Inspired Regression to create training data that works across different domains, and experiments show it outperforms existing methods.
Vision-language models (AI systems that process both images and text together) can leak private information from user-uploaded content, such as identifying people in photos or extracting sensitive text. This research examines privacy risks when users submit images and text to these models. The paper proposes privacy-preserving methods to protect user data while still allowing these AI systems to function effectively.
This survey paper examines algorithm debt in machine learning and deep learning systems, which refers to the long-term costs and problems that accumulate when developers use suboptimal algorithms or methods in AI projects. The paper defines what algorithm debt is, identifies warning signs called 'smells' that indicate its presence, and discusses future research directions. Understanding algorithm debt helps developers recognize when quick, temporary solutions in AI projects create technical problems that become harder and more expensive to fix later.
This research presents HEFLGuard, a security system that detects backdoor attacks (hidden malicious behavior injected into AI models) in homomorphic encryption-based federated learning (a privacy-protecting technique where multiple parties train an AI model together using encrypted data). HEFLGuard works by having the server create validation models from encrypted client groups and having clients compare their local model outputs on trusted samples to identify suspicious behavior, while also using Byzantine fault-tolerant aggregation (a method that functions correctly even when some participants act maliciously) to ensure robustness.
Fix: HEFLGuard constructs overlapping validation models from encrypted client groups and clients locally compare logits (output scores) of the global and validation models on benign samples to expose backdoor behavior. The framework combines consistency verification across non-IID validation groups with Byzantine fault-tolerant aggregation of client reports.
IEEE Xplore (Security & AI Journals)This paper describes a security problem in blockchain payment channels (like the Lightning Network, which allows faster transactions by bundling multiple payments together): malicious intermediate nodes can intercept funds by reading payment conditions sent in plaintext. The authors propose a solution using a new encryption method called CUAP-PRE (ciphertext unlinkable autonomous path proxy re-encryption, which encrypts payment instructions so intermediate nodes can't see or trace them) combined with an improved payment protocol that lets the final receiver control decryption rights in reverse order to unlock the funds.
Fix: The proposed solution is a secure off-chain payment protocol (SOCP) built on the new CUAP-PRE cryptographic primitive. According to the source, this protocol prevents malicious nodes by: (1) enabling the delegator to designate all trusted delegatees, (2) using ciphertext unlinkability to resist inference attacks and path tracing to ensure anonymity, and (3) implementing an enhanced multi-hop Hash Time-Lock Contract where the receiver at the end of the payment path can control decryption rights in a reversed multi-hop delegation manner to unlock the corresponding bitcoins on hold.
IEEE Xplore (Security & AI Journals)ShadowCoT is a backdoor attack (a hidden vulnerability inserted into an AI model that causes it to misbehave when triggered) that targets Chain-of-Thought reasoning, which is a technique where LLMs show their step-by-step thinking to solve complex problems. Unlike simpler attacks, ShadowCoT hijacks the model's internal reasoning process by subtly rewiring how attention flows through the model and changing intermediate representations (internal data the model creates while processing), allowing it to produce logical-sounding but harmful outputs while avoiding detection.
LVMark is a watermarking technique that hides invisible digital markers in video diffusion models (AI systems that generate realistic videos from text descriptions) to protect against unauthorized use. The method improves on existing approaches by maintaining temporal consistency (keeping the watermark stable across consecutive video frames) and preserving video quality through a specialized decoder that combines frequency analysis with color information. The technique successfully embeds 512-bit capacity watermarks that remain detectable even when videos are distorted or attacked.
Sanitizer is a new framework for federated learning (a training method where multiple computers learn together without sharing raw data) that addresses two major challenges: privacy and Byzantine attacks (when some computers send bad or malicious data). The framework uses more efficient cryptographic techniques and two new protocols for combining data from multiple sources, achieving 19-23 times faster performance than existing methods while maintaining accuracy and security.
This paper presents a new method for securely transmitting data between two users over a wireless channel where an eavesdropper might be listening. The researchers use polar coding (a technique for encoding data reliably) combined with feedback (information sent back from receiver to sender) to create correlation between transmitted signals and add dummy messages that confuse eavesdroppers. The proposed scheme achieves both reliable communication and strong secrecy without requiring simplified assumptions about the channel.
Researchers found that box-free model watermarking (a technique to protect ownership of AI models by embedding hidden marks), which is used for image processing tasks, can be defeated through black-box removal attacks (attacks where the attacker cannot see inside the protected model). The study describes three methods attackers could use to strip watermarks from protected models while keeping the images looking good, showing that current watermarking approaches are vulnerable.
Distributed Machine Learning (DML, where multiple computers train an AI model in parallel) creates challenges in determining which clients contributed what to the final model and preventing dishonest clients from falsely claiming ownership. This paper introduces MAMMON, a system using matrix commitments (a cryptographic method for verifying that data hasn't been tampered with) to verify each client's training history and establish fair ownership of the resulting model while using less computing power than existing methods.
This academic paper explores how Software Bill of Materials (SBOMs, detailed lists of all software components used in a project) can be extended to cover agentic AI systems (AI systems that can independently make decisions and take actions). The paper discusses schema extensions, how to organize and orchestrate these agentic components, and methods to evaluate whether AI systems produce reproducible results.
This research paper evaluates whether multiple AI agents working together can effectively help identify privacy threats in software systems using LINDDUN GO, a structured methodology for privacy threat modeling (a process of identifying ways a system could leak or misuse personal data). The study, published in July 2026, examines whether collaborative multi-agent LLM (large language model) systems can improve the quality and completeness of privacy threat identification compared to single AI agents or human analysis.
This academic paper proposes a dual-chain, privacy-preserving credential architecture designed to enable trust and learner agency in lifelong learning systems. The work focuses on creating secure credential management that protects learner privacy while maintaining verifiable educational records across multiple institutions and learning contexts.
This paper presents MCCENet, a deep learning system that improves biometric authentication by combining two types of palm data: palm vein patterns and palm shape. The system uses hierarchical feature fusion (a technique that exchanges information between different data types at early processing layers) and multimodal contrastive learning (a training method that helps the AI learn similar representations for related data from different sources) to better recognize individuals, achieving better accuracy than previous methods tested across eight public datasets.
This paper describes a new encryption method called FDXT that helps protect data privacy when searching encrypted files on untrusted servers. Previous methods like ODXT and SDSSE-CQ had weaknesses where attackers could leak information by analyzing search patterns and file sizes when users searched for multiple keywords together, but FDXT fixes these privacy leaks while maintaining similar or better performance.
Random prime number generators are essential for encryption and security protocols, but their output can become flawed and needs constant checking. This paper describes a machine learning approach that can validate quantum random number generators (QRNGs, devices that use quantum physics to create truly random numbers) by learning patterns in the prime numbers they produce and detecting when the output becomes biased (skewed toward certain values). The researchers tested their framework on both a quantum-based prime generator and a classical electronic noise generator, successfully identifying flawed configurations.
Face morphing attacks (blending two faces together to fool facial recognition systems) threaten security systems used at borders and for digital identity checks, and detecting them from a single image is difficult because there's no trusted reference image to compare against. This paper presents R-FLoRA, a new detection method that combines high-frequency image analysis (looking at fine details) with a frozen, large-scale vision transformer (a type of AI model trained on images) to spot morphing artifacts while keeping the overall understanding of the face intact. The method outperforms nine other detection approaches on multiple test datasets and works efficiently in real-world biometric verification systems.
SRAM PUFs (physically unclonable functions, which are hardware features that generate unique secret keys from a chip's manufacturing variations) suffer from reliability problems because bits can flip and change values unpredictably. This paper introduces TMVS (Threshold-based Majority Voting Scheme), a software-based method that reduces noise and fixes bias issues in SRAM PUFs while keeping the approach simple and avoiding the complexity of heavy error-correction codes.