Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Large language models (LLMs, which are AI systems trained on vast amounts of text) are vulnerable to serious attacks like hallucinations (making up false information), jailbreaks (tricking the AI into ignoring its safety rules), and backdoors (hidden malicious instructions inserted during training). This research proposes a detection method using hidden state forensics (analyzing the internal numerical patterns that flow through the model's layers) to identify abnormal or malicious behavior in real-time, achieving over 95% accuracy with minimal computational cost.
HKT-SmartAudit is a framework that creates smaller, faster AI models specifically trained to find bugs in smart contracts (self-executing code on blockchain networks). The framework uses knowledge distillation (a technique where a large, accurate AI model teaches a smaller model by sharing what it has learned), allowing these lightweight models to detect vulnerabilities effectively while using far less computing power than larger models.
Wireless sensing uses Wi-Fi and similar signals to detect human activity like movement and sleep patterns, but broadcast signals can be intercepted by unauthorized users, creating privacy risks. This research proposes using multi-antenna signal processing (techniques that use multiple receiving antennas to manipulate wireless signals) as a privacy protection method at the physical layer (the lowest level of wireless communication, before encryption). The study analyzes the tradeoff between sensing accuracy and privacy protection by modeling the system mathematically and defining performance boundaries.
This paper presents CIBPU, a new secure branch prediction unit (BPU, a component that helps processors predict which instructions to execute next) that protects against attacks trying to infer sensitive information by observing how the BPU behaves. Unlike previous designs that either isolated the BPU physically or used encryption with frequent key updates, CIBPU uses redundant storage (extra copies of data), smart indexing, and encryption without periodic key changes to hide branch conflicts (situations where different instructions compete for the same storage space) from attackers. The researchers tested CIBPU in simulators and on real hardware, finding it adds only about 2-4% performance slowdown, which is better than other secure branch prediction approaches.
LitCVit is a lightweight AI model designed to detect malicious encrypted network traffic (data sent over secure connections) without needing to decrypt it or manually extract features. The model uses self-supervised learning (training where the AI learns patterns from unlabeled data) and vision transformers (a type of neural network architecture) to analyze patterns across multiple data packets and flows (sequences of related network communications) while running much faster than existing approaches, achieving 98% accuracy on test datasets.
This paper describes a vulnerability in 5G networks where an attacker can intercept scheduling information from downlink control information (DCI, the signals that tell devices which radio resources to use) and use it to jam the PUSCH (physical uplink shared channel, the main data transmission channel from devices to the network). To defend against this DCI sniffing-based smart jamming attack, the researchers propose a suppression method that identifies which DCI-scheduled resources are being attacked and reconstructs the PUSCH transmission by leveraging differences in spatial domain features between legitimate users and attackers.
Fix: The proposed suppression method leverages DCI-scheduled subset identification and PUSCH resource reconstruction. It fundamentally relies on differences in spatial domain features under available control channel elements and resource block group granularities between legitimate users and the attacker, to selectively exclude unwanted elements while safeguarding the authenticity of targeted transmissions.
IEEE Xplore (Security & AI Journals)Federated learning (FL, where multiple devices train AI models together without sharing raw data) faces privacy risks because adversaries can extract sensitive information from model updates. FedNSA is a new protocol that combines differential privacy (adding mathematical noise to hide individual data patterns), encryption, and multi-party computation (MPC, a technique where multiple parties jointly compute results without revealing their individual inputs) to protect model updates while reducing the communication and computational burden that makes secure aggregation impractical on resource-constrained devices like smartphones.
FALCON-Net is a detection system designed to identify AI-generated images by analyzing their technical flaws. The system works by examining two key weaknesses in generated images: the lack of device-specific sensor noise (natural imperfections that real cameras add) and unnatural pixel intensity variations that result from oversimplified generation processes. FALCON-Net combines two analysis modules (one for noise patterns and one for local pixel variations) to reliably distinguish AI-generated images from real ones, even when tested on image generation models it wasn't trained on.
Screaming channels are a type of side-channel attack (a method of stealing data by analyzing electromagnetic leakage from a device) that can work from several meters away on devices with both wireless (RF) and digital components on the same chip. This research shows that attackers can find useful leakage signals at many more frequencies than previously thought, not just at the harmonics (multiples) of the clock frequency, making attacks more effective even in noisy RF environments and at greater distances.
Privacy labels on app stores like Google Play are supposed to give users a quick summary of what data apps collect, but many developers fail to accurately report their practices. Researchers created PriLabel, a tool that analyzes decompiled code (machine-readable instructions extracted from apps) to automatically detect when apps transmit sensitive data without disclosing it in their privacy labels. Testing on thousands of popular apps found that many failed to report data collection, including some that transmitted financial information like credit card numbers without proper labeling.
This paper presents a new method for task-adapted compressed sensing MRI (CS-MRI, a technique that reconstructs medical images from fewer measurements than normally required). The method uses information theory (a branch of mathematics about data and uncertainty) to improve both the accuracy of image reconstruction and the reliability of clinical diagnoses, while also allowing flexible control over how many measurements are needed. The approach can handle different clinical scenarios, including one that suppresses image reconstruction to protect patient privacy.
This research paper, published in June 2026, explores how to make multi-party private set union (a process where multiple parties combine datasets while keeping their individual data secret) more efficient and secure against malicious attacks. The authors propose using trusted execution environments (TEEs, hardware that protects code and data even from the computer's owner) to achieve this goal. The paper aims to balance computational efficiency with strong security guarantees when multiple parties need to collaborate while protecting sensitive information.
Researchers used LLMs (large language models, AI systems trained on vast text data) and model checking (a technique to verify if software behaves correctly by examining all possible states) to study Android's permission system, which controls what apps can access on your phone. The study involved modeling how this system works, checking if it's secure, and finding ways to exploit it using AI techniques.
This research paper examines how visual explanation techniques can help protect federated learning (a machine learning approach where multiple computers train a model together without sharing raw data) from poisoning attacks (attempts to corrupt the training data or model). The authors propose an enhanced version of LayerCAM (a method that visualizes which parts of an input an AI focuses on), combined with autoencoders (neural networks that compress and reconstruct data), to detect and defend against such attacks.
AICCE is an AI system designed to automatically check whether network communication follows IPv6 protocol standards (the rules for how internet data should be formatted and transmitted in the newer IPv6 system). The system uses retrieval-augmented generation (RAG, where an AI pulls in relevant protocol specifications to answer questions) and multiple AI agents that debate decisions to catch both obvious and hidden violations that traditional rule-based checkers miss.
This paper addresses challenges in Online Continual Learning (OCL, a type of AI training where a model learns from streaming data with unknown task boundaries) by proposing Quadruplet Augmentation, a method that uses four augmentation strategies to preserve two key properties: attribute invariance (keeping object characteristics consistent across learning sessions) and structure invariance (maintaining relationships between different attributes). The approach uses techniques from Fourier analysis (mathematical transformation of signals) and channel independence constraints to improve how AI models learn from continuously arriving data without forgetting previously learned information.
This paper addresses node classification in hypergraphs (a type of graph where edges can connect more than two nodes) by proposing a method called LHCCA that learns from labeled data in a source hypergraph to help classify unlabeled nodes in a target hypergraph. The method improves on existing approaches by considering both local relationships (direct connections) and high-order relationships (connections at greater distances), combining these through an attention mechanism (a technique that learns which parts of the input to focus on), and using adversarial domain adaptation (a training strategy to make learned features work across different hypergraphs) and contrastive learning (a method that learns by comparing similar and dissimilar examples).
Researchers developed new data augmentation strategies (techniques for artificially expanding training datasets) to improve website fingerprinting, which is a method to identify which websites users visit by analyzing their network traffic patterns. The study, published in August 2026, demonstrates how combining multiple augmentation approaches can make these fingerprinting techniques more effective.
This research paper describes a method called ReSLC that protects AI systems used to find software bugs from backdoor attacks, where attackers secretly embed malicious instructions into the AI's training process. The approach uses redundant semantic LLM compression (a technique that removes unnecessary information from large language models while keeping their core abilities) to make these hidden attacks harder to carry out. The work was published in July 2026 in the Journal of Information Security and Applications.
This research paper presents a hybrid deep learning method using autoencoders (neural networks that learn to compress and reconstruct data) and transformers (AI models that process sequences of information) to detect a new type of attack called unresponsive ECN attacks on low-latency network services (systems designed to minimize delay in data transmission). The proposed method achieves over 90% accuracy in detecting these attacks while keeping false alarms below 0.01%, outperforming existing detection approaches by more than 10%.