Research

This paper introduces SENSA (Self-Evolving Neuro-Symbolic Autonomy), a system that combines neural learning (data-driven pattern recognition) with symbolic reasoning (rule-based logic) to help AI make better decisions in complex, changing environments. The system uses causal reasoning (understanding why things happen, not just what happens) and can automatically improve its decision-making policies while staying robust against adversarial attacks (attempts to trick or disrupt the AI).

Researchers discovered that SemAder, a technique that manipulates both the structure and meaning of binary code (compiled programs), can fool LLM-based binary code analysis tools into missing security problems. The study shows that by carefully modifying how code is organized and what it semantically does, attackers can evade detection systems that use large language models to analyze compiled programs for vulnerabilities.

This release document describes updates to security techniques and mitigations related to AI systems, including new attack methods like stealing web session cookies and jailbreaking LLMs (large language models, AI systems trained on massive amounts of text), as well as updated defenses like AI guardrails (safety features that prevent unwanted outputs) and telemetry logging (monitoring system activity). Several new case studies illustrate real-world attacks on AI services, including prompt injection (tricking an AI by hiding malicious instructions in user input) and vulnerabilities in commercial AI products.

This research review examines how the metaverse (immersive virtual environments powered by AR, VR, and AI) is changing digital marketing while creating ethical concerns around privacy, transparency, authenticity, and fairness. The authors analyzed 50 academic reviews and found that addressing these ethical challenges requires coordination between marketers, platform companies, policymakers, and users.

This research presents EthicAR, a Safe Reinforcement Learning (Safe RL, a training method that teaches AI systems to make decisions while avoiding harmful outcomes) framework for autonomous vehicles that prioritizes protecting vulnerable road users like pedestrians and cyclists. The system uses ethics-aware cost signals (penalty measures that make the AI weigh safety and harm severity) combined with a special learning technique called Temporal Cost Aggregation (TCA, which tracks risks across multiple decision steps) to reduce collisions by 20-45% compared to standard methods. The approach combines formal control theory with machine learning to create autonomous driving behavior that is both safer and more ethically accountable.

Federated learning (FL, a technique where AI models are trained across multiple computers without sending raw data to a central server) is vulnerable to attacks where dishonest participants send corrupted model updates that poison the final model without being detected. This paper introduces AMBER, a framework that adds three layers of verification to check whether clients are trustworthy: confirming data hasn't been tampered with, detecting when clients provide misleading inputs, and verifying that model computations are correct using a trusted execution environment (TEE, a secure area of a computer processor that protects sensitive operations).

This research addresses the challenge of detecting deepfakes (fake videos created by AI) that use multiple types of data like video and audio together, as these deepfake techniques become more advanced. The researchers propose Amber, a system that uses continual learning (where an AI improves over time as it sees new examples) to better remember what real deepfakes look like while adapting to new deepfake methods, inspired by how the immune system selects useful cells.

HookSteg is a new steganography method (hiding secret data inside images) that achieves zero-error transmission of hidden information through lossy channels like social media platforms. It uses process hooking (intercepting system function calls to modify behavior) and dynamic taint tracking (monitoring how data is altered) to detect and bypass compression and scaling attacks that normally corrupt hidden data, improving extraction accuracy by 36.63% compared to existing methods.

This academic survey article examines agentic AI in healthcare, which refers to AI systems that can independently plan and execute tasks to accomplish goals. The article discusses both the potential benefits of using such AI systems in medical settings and the technical, ethical, and practical obstacles that need to be addressed. The survey provides an overview of current research directions for developing safer and more effective autonomous AI agents in healthcare applications.

This research addresses video-based visible-infrared person re-identification (VVI-ReID, a technique that identifies the same person across visible light and thermal infrared video footage for surveillance) using unsupervised learning, which doesn't require expensive labeled training data. The authors propose Causal Bootstrapped Alignment (CBA), a framework that uses temporal video information and causal intervention (a method that identifies cause-and-effect relationships by simulating changes) to improve how well the system recognizes people across both imaging modes.

Representation engineering (RepE, a technique that modifies how an AI model processes information internally) defenses against jailbreaks (methods that trick LLMs into ignoring safety rules) have relied on block-list approaches that learn from known harmful examples, but this leaves gaps against new attacks. The paper proposes Ellipsoid Control, a white-list defense that instead protects the model's normal, harmless behavior by using test-time defense (constraints applied when the model runs) with projected gradient descent (a mathematical technique that adjusts the model's internal state while keeping it close to safe boundaries) to trigger refusals on harmful inputs while preserving the model's ability to answer legitimate questions.

This research proposes PRBPR, a privacy-preserving redactable blockchain system that addresses conflicts between blockchain immutability and regulations like GDPR's right to be forgotten. The scheme combines hierarchical blinding factors (mathematical techniques for hiding information), chameleon hash trapdoors (special keys that allow controlled modification of blockchain records), and ciphertext-policy attribute-based encryption (encryption that grants access based on user attributes while hiding who can access data) to enable efficient data deletion, fine-grained access control, and auditability while maintaining security in multi-party data sharing.

This research addresses two problems in federated learning (a training approach where multiple computers work together to improve a shared AI model without sending raw data to one central location): client drift (when different computers' local updates diverge from each other due to unequal data) and Byzantine attacks (when some computers send false or malicious updates to sabotage the model). The paper proposes two frameworks called DRAG and BR-DRAG that use a metric called divergence of degree to measure how far off-track each computer's updates are, then automatically correct them through a process called linear calibration, with BR-DRAG adding a vetted dataset at the main server to defend against malicious actors.

This research presents a new algorithm for training predictive models on sensitive data while protecting privacy using differential privacy (DP, a mathematical technique that adds noise to data to prevent identifying individuals). The algorithm extends previous privacy-preserving methods to handle weighted empirical risk minimization (wERM, where different data points contribute differently to model training), which is particularly useful for personalized medical treatment decisions. Testing shows the approach successfully protects privacy while keeping the trained models effective.

Multimodal Online Federated Learning (MMO-FL, a system where multiple IoT devices with different types of sensors train AI models together in real-time without sending raw data to a central server) faces security risks from data poisoning attacks (when attackers inject corrupted or malicious data into the training process). Researchers studied these attacks across three vulnerability dimensions (distributed systems, real-time learning, and multiple data types) and proposed a detection and mitigation algorithm tested on real-world datasets to defend against them.

This research presents a new protocol called Updatable MPSI (UMPSI) that helps multiple organizations securely share threat intelligence data while keeping their information private. Unlike existing systems that require expensive full re-computation when threat data changes, UMPSI uses lightweight mechanisms to quickly update only the new or changed threat indicators, making real-time collaborative threat detection practical.

Runtime side-channel attacks exploit timing differences in program execution to steal secrets like encryption keys or sensitive data. Most existing defenses try to make code execution times equal across different paths, but they don't account for the specific hardware device the program runs on, which can cause problems like unnecessary slowdowns or incomplete fixes. DISARM is a new approach that combines hardware and software knowledge by measuring actual timing on real embedded devices (small, specialized computers) to create better, more targeted fixes that work correctly across different hardware while reducing performance overhead.

This paper presents a new method for rotation estimation (determining how an object is oriented in 3D space), which is important for computer vision and robotics, especially in safety-critical applications. Instead of treating rotation estimation as a difficult non-linear optimization problem, the researchers reformulated it as a linear model fitting problem and developed a voting-based approach that can handle large amounts of noise and corrupted data very quickly using GPUs (graphics processing units, which are good at parallel computing).

LoRASculpt+ is a framework that improves how multimodal large language models (MLLMs, AI systems that process both text and images) can be adapted to specific tasks while keeping their general knowledge. The method uses sparse LoRA updates (a parameter-efficient technique that only modifies a small set of weights rather than all parameters) with regularization to reduce unnecessary changes during training and prevent the model from forgetting what it already knows.