Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Researchers discovered a new threat called ToLaR (threat of loads and renewables) that can attack modern power grids with renewable energy sources like solar panels and batteries more efficiently than previous attacks. By manipulating both the electricity demand side and the renewable energy generation side simultaneously, attackers can cause grid instability with only 10% of the resources needed for older attack methods, potentially dropping grid frequency to dangerous levels.
This research proposes TAPGuard, a framework for detecting cascading threats in Trigger-Action Programming (TAP, a system where one event automatically triggers another action, commonly used in smart home devices). The framework uses large language models (AI systems trained on text) to understand the semantic meaning (the actual intent and meaning, not just the structure) of automation rules and identifies two types of threats: explicit ones from direct device interactions and implicit ones from rules sharing environmental variables that shouldn't interact. TAPGuard performs better than existing methods at catching these dangerous rule combinations.
Researchers created SOOM, a defense method that obfuscates (hides or disguises) deep learning operators to protect against model extraction attacks, where attackers reverse-engineer compiled neural network code to recreate trainable models. Built on TVM (a deep learning compiler), SOOM uses a machine learning cost model to scramble how operators work while keeping inference fast, achieving a 89% failure rate against extraction attacks with minimal performance slowdown.
Fix: The source proposes SOOM itself as the mitigation: a schedule-search-based operator obfuscation method built on TVM that constructs an obfuscation space for deep learning operators and uses a security-aware learned cost model based on XGBoost gradient boosted trees to generate obfuscated executable code for various deep learning operators, balancing security objectives with performance requirements.
IEEE Xplore (Security & AI Journals)Existing model fingerprinting techniques (methods that create unique digital signatures to prove ownership of AI models) are vulnerable to false claim attacks, where attackers can fraudulently claim they own models they didn't create. This paper introduces FIT-Print, a targeted fingerprinting approach that uses optimization to create verifiable signatures resistant to these false claims, offering two specific methods (bit-wise FIT-ModelDiff and list-wise FIT-LIME) that achieved 100% success in preventing false ownership claims while maintaining accurate ownership verification.
Fix: The paper proposes FIT-Print, a targeted fingerprinting paradigm that 'actively counters false claim attacks' by leveraging 'optimization to transform the fingerprint into a verifiable, targeted signature.' Two specific black-box fingerprinting methods are introduced: 'bit-wise FIT-ModelDiff' which 'utilizes output distances' and 'list-wise FIT-LIME' which utilizes 'feature attributions as robust model signatures.' The framework demonstrated '100% defense success rate' against false claim attacks and '100% ownership verification rate.'
IEEE Xplore (Security & AI Journals)Text-to-image diffusion models (AI systems that generate images from text descriptions) can be misused to create fake or harmful images, and current defenses using hidden noise patterns don't work well enough. Researchers propose DANP (Dual Attention-Guided Noise Perturbation), a defense method that adds imperceptible perturbations (tiny, invisible changes) to confuse the model's understanding by manipulating how it pays attention to different parts of the image and how it predicts noise during generation.
Fix: The proposed solution is the Dual Attention-Guided Noise Perturbation (DANP) immunization method, which works by: adding imperceptible perturbations across multiple timesteps, using dynamic thresholds to identify text-relevant and irrelevant regions, reducing attention in relevant areas while increasing it in irrelevant ones to misdirect malicious edits, and maximizing the discrepancy between injected noise and the model's predicted noise to further interfere with generation. The authors state this method achieves state-of-the-art performance against malicious edits.
IEEE Xplore (Security & AI Journals)MEC-Dedup is a security approach for mobile users storing data in cloud systems that use edge computing (processing done on devices near the user rather than in distant data centers). The system addresses risks that arise when multiple users' identical files are deduplicated (combined into one copy to save space), which could let attackers identify sensitive information. The research proposes methods to keep user data secure while still allowing the efficiency gains of deduplication in edge-assisted cloud storage.
This academic paper presents a new cryptographic method for secure communication between vehicles and infrastructure in VANETs (vehicular ad hoc networks, which are temporary networks formed by moving vehicles). The scheme uses identity-based aggregate signcryption (a technique that combines digital signatures for authentication with encryption for confidentiality, while processing multiple messages together), and the authors claim to have mathematically proven it cannot be broken by attackers.
This academic paper presents a new method for detecting unusual network activity using parallel GANs (generative adversarial networks, AI systems that learn patterns by comparing real data against artificially generated data) without requiring manually set detection thresholds (cutoff points that decide what counts as suspicious). The approach uses comparative reconstruction error learning, meaning it compares how well the AI can recreate normal network behavior to spot deviations that might indicate attacks or intrusions.
This research article examines whether training employees on cybersecurity improves their understanding of password security and safe Wi-Fi practices. The study, published in June 2026, investigates the connection between formal security education and employees' actual knowledge of protective measures in their daily work.
This research paper analyzes how networked control systems (computer systems where multiple sensors share information across a network) behave when under attack from false data injection, or FDI (inserting fake sensor readings to disrupt the system). Using game theory (a mathematical framework for analyzing competing strategies), the researchers model the conflict between legitimate system operators trying to keep data accurate and attackers trying to corrupt it, then prove that both sides will reach a stable strategic equilibrium (a predictable outcome where neither side can improve by changing tactics alone).
This paper introduces MFA-DPRF, a new multi-factor authentication (MFA, a security method requiring multiple forms of proof like a password and a code) scheme designed to handle two problems that existing systems ignore: password recovery when users forget their login credentials, and fine-grained access control (letting administrators set specific rules about which users can access which resources). The system works by requiring users to provide both a valid password and attributes that match an access policy, and it includes a dynamic password recovery method using secret questions and secret sharing (splitting a secret into pieces so no single piece reveals the secret).
Researchers developed a new membership inference attack (MIA, a method to determine whether specific data was used to train an AI model) called MU-MIA that uses machine unlearning (a technique to make a model forget specific training samples) to track how a model forgets information about individual samples. The attack works by monitoring changes in the model's behavior as it unlearns each sample and uses a BiLSTM classifier (a type of neural network that analyzes sequences of data) to distinguish between samples that were in the training data versus those that weren't.
A study in Sierra Leone tested whether AI (specifically Google's Gemini) could help students learn math better by acting as a teaching partner rather than replacing teachers. The AI was designed using a 'Socratic' approach, asking guiding questions instead of giving direct answers, and students who used it showed significant learning gains equivalent to 1.2 to 2.5 years of typical progress in just eight weeks, while maintaining high engagement and shifting their own questions toward understanding rather than just seeking solutions.
This academic paper describes a new encryption method designed to let multiple people search through encrypted medical data while protecting privacy and controlling who has access. The scheme uses lattice-based cryptography (a type of math-hard encryption based on complex grid structures) and allows for selective disclosure (sharing only certain information with specific people) and user revocation (removing someone's access rights). This addresses the challenge of keeping medical information secure while still making it searchable and shareable in healthcare systems.
This academic paper describes a method for securely sharing secret images among multiple people using the Chinese Remainder Theorem (a mathematical technique for solving certain types of equations) and polynomials (mathematical expressions with variables). The scheme includes a certification process to verify that the shared image pieces are authentic and haven't been tampered with.
This academic paper proposes TMAS, a threshold multi-auditor auditing scheme designed to verify data integrity and security in cloud-fog computing environments (distributed systems where data processing happens both in the cloud and at edge devices closer to users) where trust between parties is limited. The scheme uses multiple independent auditors working together so that no single auditor needs to be completely trusted, addressing the challenge of maintaining security when collaborating systems don't fully trust each other.
STAFF is a research tool for testing firmware (the low-level software that runs on hardware devices) by using fuzzing (automated testing that feeds random or specially crafted inputs to find bugs). The tool uses stateful taint analysis (tracking how untrusted data flows through a program) to improve the fuzzing process and find security vulnerabilities more effectively in full systems.
Researchers have developed PLC-Defuser, a tool that detects hidden malicious code (logic bombs, which are programmed instructions designed to execute harmful actions when triggered) in PLCs (programmable logic controllers, computers used to automate industrial equipment like factory machinery). The tool uses control flow graphs (visual maps showing how a program's instructions connect and execute) and model checking (automated verification that tests whether software meets safety properties) to find these threats before they can cause damage.
This research paper describes a method for creating attack graphs (visual maps showing how attackers could move through a company's computer network) by using data collected from endpoint devices (individual computers and servers). The approach is designed to scale efficiently, meaning it can handle large enterprise networks without becoming too slow or resource-intensive. The work was published in October 2026 in the journal Computers & Security.
CoolTest is a new randomness test designed to work well with small amounts of data, published in November 2026. Randomness tests check whether data appears truly random or follows a pattern, which is important for security applications like cryptography (the practice of encoding information to keep it secret). This tool addresses a limitation of existing tests that often require large datasets to work accurately.