Browse All

AI agents themselves cannot be held legally responsible for damage they cause, but the people and companies that deploy them (put them into use) can be held liable instead. Experts warn that deployers should take responsibility for foreseeable harms their AI agents might cause, even if the harm wasn't intentional.

SymSyz is a hybrid kernel fuzzing framework that combines kernel fuzzing (automated testing that sends random inputs to find bugs) with symbolic execution (a technique that analyzes code paths mathematically to reach difficult-to-access code). The framework addresses three main problems: inconsistent states between the fuzzer and symbolic engine, overhead from tracking too much symbolic information, and slow constraint solving. SymSyz improves testing speed and bug detection by using snapshot-based environment synchronization, limiting symbolic tracking to specific code origins, and prioritizing constraint solving.

Text-to-image models (AI systems that generate pictures from text descriptions) can be tricked by attackers who fine-tune them (adjust their parameters on new data) to bypass safety protections and create unsafe images. This paper introduces Patronus, a defensive framework that makes these models more resistant to such attacks by using a specially trained safety decoder (a component that processes the model's internal representations) that produces corrupted outputs for unsafe content while preserving normal image generation for safe requests.

This research proposes a new method for cloud storage that combines data deduplication (removing duplicate copies of files) with integrity auditing (verifying data hasn't been corrupted or altered) while protecting user privacy. The key innovation is using lightweight algebraic operations (simple mathematical calculations) instead of expensive BLS signatures (complex cryptographic signatures), and applying randomized blinding techniques (adding random noise to hide identifying information) to prevent auditors from discovering who owns which files.

Machine unlearning (MU) is a technique that removes the influence of specific data from trained AI models without retraining them from scratch, which is important for privacy laws like the right to be forgotten. DeepU is a new framework that performs fine-grained unlearning by analyzing individual weights (the parameters that make up a neural network) within each layer and categorizing them as influential, intra-dependent, or non-influential, then applying targeted updates like resetting or adjusting those weights. Tests show DeepU reduces successful membership inference attacks (where attackers try to figure out if specific data was used in training) by 60-90% while dropping accuracy by less than 3%, and it's significantly faster than other methods.

This paper introduces LICA, a new security method that can detect Incorrect-Plaintext Attacks (IPA, where someone encrypts wrong data or falsely claims decryption failed) even against adversaries with quantum computers. LICA uses lattice-based cryptography (mathematical problems believed to be hard even for quantum computers) combined with commitments (binding agreements on encryption parameters) to identify which party is acting maliciously, and presents LICADS, a data sharing protocol that catches quantum IPA by having a trusted arbitrator challenge suspicious parties and assign penalties to the dishonest one.

Federated learning (FL, a way for multiple computers to train an AI model together without sharing raw data) faces a tradeoff between privacy and performance: encryption methods are slow, while differential privacy (DP, adding noise to data to hide individual information) reduces accuracy. This research proposes a cascaded filter that selectively adds noise only to the most important model parameters (the dimensions with large values and high variation) before sending them to a central server, achieving both privacy protection and better model performance than existing methods.

Distributed Learning (DL, training AI models across multiple computers) faces problems when some computers are slow (stragglers) or malicious (Byzantine nodes, computers that send incorrect data). FVCC is a new framework that uses coded computing (a technique that adds redundancy so missing data can be recovered) with faster decoding and verification methods to make distributed learning more robust and efficient. The system uses a bidirectional two-dimensional ZigZag Decoding algorithm to recover data quickly and Freivalds' algorithm (a lightweight verification method) to detect dishonest computers.

Zero-shot voice conversion (ZSVC, technology that can change a person's voice characteristics without training data) creates risks of audio forgery and copyright violations. Researchers propose VocaLock, a watermark-based detection system (invisible digital markers embedded in audio) that can both identify forged audio and identify whose voice was stolen when ZSVC is used maliciously. VocaLock embeds watermarks in the audio spectrum (frequency information) in a way that survives voice conversion attacks while maintaining audio quality.

MDIGuess is an AI system that improves targeted password guessing attacks by combining multiple types of information about a user, including their past passwords, personal details (PII, or personally identifiable information), and passwords leaked from other services. The system uses an autoregressive neural framework (a type of AI that predicts passwords one piece at a time based on patterns) and achieves a 35% success rate cracking passwords within 1,000 guesses, significantly outperforming existing attack methods.

GeoPrivd is a framework for analyzing traffic video data while protecting people's privacy through differential privacy (a mathematical technique that adds noise to data to hide individual information). Instead of storing detailed tracking information about people's movements, GeoPrivd uses a query language called GeoPrivdQL that lets analysts ask questions about traffic patterns while automatically ensuring privacy is preserved without exposing sensitive trajectory data.

ESecDT is a framework that allows multiple parties to train decision trees (machine learning models used to make predictions by sorting data into categories) together while keeping their individual data private. It combines two cryptographic techniques called Function Secret Sharing (FSS, a method where a secret is split into parts that only work together) and Replicated Secret Sharing (RSS, another way to distribute secrets across parties) to reduce the amount of data that must be sent between parties during training while maintaining strong privacy protections.

Researchers developed a new detection system for AI-generated images (AIGC, or AI-generated content) that combines three techniques to identify fake images created by generative models (AI systems that can create new images from scratch). The system works by examining noise patterns at the sensor level, identifying unusual patterns in the frequency domain (how colors and patterns repeat), and highlighting suspicious regions in images, allowing it to detect synthetic images even when faced with new or unknown generative models.

DiEL is a method for improving face images while keeping the person's identity recognizable, especially in difficult conditions like extreme angles, blurriness, or poor lighting. The approach uses evolutionary learning (a technique that evolves solutions over time) to separate identity information from pose information (head angle), then reconstructs faces using a pose dictionary (a library of standard face angles learned from many images) to maintain consistency. The method outperforms existing approaches by an average of 4.66% on six benchmark datasets, with particularly strong improvements on challenging cross-pose tests.

This research describes a new attack method called TDOA (Textual Dynamic Outputs Attack) that can trick large language models by exploiting a weakness in how they handle variable outputs. Unlike older attack methods that assume a fixed set of possible answers, real-world LLMs often generate answers that go beyond predefined categories or produce different numbers of labels depending on the input, creating what researchers call 'dynamic outputs.' TDOA works by using a clustering approach (grouping similar outputs together) to convert these unpredictable outputs into a simpler form that existing attack techniques can target, achieving up to 80.8% success rates with very few queries.

This research proposes a secure system for multimodal biometric recognition (using multiple biological measurements like palmprints and vein patterns together) that protects user privacy by creating cancellable templates (protected versions of biometric data that can't be reversed to recover the original). The system uses a neural network to combine features from different biometric sources and a novel random projection method to generate templates that are irreversible, revocable (can be replaced if compromised), and resistant to various attacks.

Existing AI systems for finding bugs in smart contracts (programs that run on blockchains) struggle because they only learn from known vulnerabilities and use limited types of information about the code. This paper presents Synesthete, a new detection method that combines multiple types of code features (text, graph structures, and images from different code representations) and uses self-learning to better identify both known and previously unseen vulnerabilities in smart contracts.

This research presents a new authentication protocol (a method for verifying identity between two parties) designed for Vehicle-to-Grid (V2G) communication, which allows electric vehicles to securely exchange energy with charging stations. The protocol uses lattice-based cryptography (a type of math-based encryption resistant to attacks from quantum computers, which are extremely powerful future computers) and was tested for security using both automated tools and manual analysis to ensure it works safely in real-world situations.