Research

PadNet is a defense method designed to protect neural networks (AI models that learn patterns from data) against adversarial examples (specially crafted inputs that trick AI systems into making wrong predictions). The paper, published in an academic journal, presents techniques to make these AI systems more robust when facing such attacks.

Deep neural networks (machine learning models with many layers that process information) are vulnerable to adversarial examples, which are inputs slightly modified to fool the AI into making wrong predictions. This paper proposes adding a redundant fully connected layer (a type of neural network component that connects all inputs to all outputs) with a special loss function to make these networks more robust against attacks while maintaining accuracy on normal inputs.

This research presents a new framework called Propose-Rectify that helps detect and locate image manipulations (alterations made to photos) by combining two approaches: first, a semantic reasoning stage uses a modified LLaVA model (a multimodal AI that understands both images and language) to identify suspicious regions, and second, a refinement stage uses specialized forensic analysis (technical methods that detect tampering traces) to validate and precisely locate the manipulated areas. The framework bridges the gap between AI understanding and forensic detection, achieving better accuracy than previous methods.

Semi-supervised learning (SSL, a training method where models learn from both labeled and unlabeled data) is vulnerable to backdoor attacks, where attackers can corrupt model predictions by poisoning a small portion of training data with hidden triggers. This paper reveals that SSL backdoor attacks are particularly dangerous because they exploit the pseudo-labeling mechanism (the process where the model assigns labels to unlabeled data) to create stronger trigger-target correlations than in supervised learning. The researchers propose Backdoor Invalidator (BI), a defense framework using complementary learning, trigger mix-up, and dual domain filtering to obstruct and filter backdoor influences during both feature learning and data processing.

Connected autonomous vehicles (CAVs) use multiple types of sensors, like LiDAR (light-based radar that creates 3D maps) and cameras, to understand their surroundings, and combining information from both sensors improves accuracy. However, this sensor fusion process can leak private information and relies on a third party to generate random numbers, which could be compromised by attackers. Researchers propose MPOF, a model that uses secure computation protocols (mathematical methods that let systems calculate results without exposing raw data) and sacrificial verification (a technique that detects when a third party behaves maliciously) to protect privacy while defending against attacks from that third party.

This article proposes two types of neural networks (machine learning models inspired by how brains work) designed to solve resource allocation problems (deciding how to distribute limited resources) in networked systems like smart grids. The neural networks are built to resist disturbances (unwanted interference or noise) by using different strategies: one exploits known patterns in the system, while the other uses an observer (a component that monitors the system state) to handle unexpected interference. Both networks were mathematically proven to work correctly and tested with simulations.

This research proposes a new algorithm for distributed aggregative optimization (a problem where multiple agents must work together to optimize a goal that depends on all their decisions). To protect privacy during information sharing between agents, the algorithm adds Laplace noise (random mathematical perturbations) to exchanged data and uses a noise deduction mechanism (a technique to prevent errors from building up due to the noise), achieving what researchers call differential privacy (a formal guarantee that individual data cannot be easily identified from the output).

Researchers found that modern natural language processing (NLP) models, which are AI systems trained to understand text, are very vulnerable to "rubbish text" attacks where sentences are heavily modified to become nonsensical to humans but still produce the same prediction from the model. The team developed a new algorithm called MOSA-S2 that uses multiobjective simulated annealing (a optimization technique that balances multiple competing goals) and stopword substitution (replacing words with meaningless filler words) to generate better adversarial examples, revealing that these NLP models may not truly understand language semantics despite their confidence in predictions.

Researchers are using GANs (generative adversarial networks, AI systems that create synthetic data by having two neural networks compete against each other) to generate sophisticated adversarial attacks that can fool IDSs (intrusion detection systems, software that monitors networks for suspicious activity). This review examines how GAN-based attacks compromise IDS security and identifies which GAN variants are most effective at creating realistic attacks that evade detection.

Federated learning (FL, a system where multiple participants train AI models on their own data and share only the updates rather than raw data) is vulnerable to targeted attacks where malicious participants send bad updates to corrupt the shared model. This paper proposes FedDOT, a defense framework that uses two graph-based detection algorithms (maximum spanning tree-based attacker detection and densest graph-based attacker detection) to identify and filter out malicious updates, even when attackers make up more than half of the participants.

TrustAI is a methodology for evaluating whether AI systems are trustworthy, addressing the current lack of standard ways to certify AI reliability. The approach breaks down trust into high-level requirements (called pillars of trust) and then into specific questions that AI experts answer, using fuzzy logic (a technique for handling uncertain or imprecise information) to combine their judgments into a trust score and generate certificates for systems that meet standards.

TrustTiny-HAR is a system that enables microcontrollers (small, low-power computers) to recognize human activities while being honest about uncertainty, rather than confidently guessing when faced with unfamiliar situations. The system combines a compact machine learning model with special techniques to detect when it lacks sufficient evidence to make a prediction and can optionally send only a small data summary (a 32-64 dimensional feature sketch) to a more powerful device instead of raw sensor data. When tested on real activity datasets with various challenging conditions like unseen activity types and sensor placement shifts, it maintained strong recognition accuracy while improving calibration (confidence reliability) and reliably rejecting out-of-distribution scenarios.

This research paper presents CLIP-ADA, a method for detecting synthetic images (fake images created by AI generators) that works better across different types of generators and artifacts. The method analyzes how CLIP (a vision-language model that understands both images and text) processes images at different levels, then uses this understanding to train detectors that rely less on specific artifact patterns and more on general forensic features, achieving over 6% better accuracy on unseen synthetic images.

Researchers developed SRAP (Self-Reversible Adversarial Patch), a technique that creates adversarial patches (small, intentionally corrupted image regions designed to fool AI models) that can be reversed back to the original image while protecting privacy. The method improves two key weaknesses in existing adversarial patches: transferability (working across different AI models, achieving up to 90% success rate) and robustness (resisting image processing and defensive techniques), and demonstrates an 88% attack success rate against commercial AI services.

This paper presents TSFA, a method for handling unsupervised open-set domain adaptation (UOSDA, a machine learning challenge where an AI model must work with data from a new environment that may contain classes it wasn't trained on) in time-series classification. The method uses a two-stage approach: first extracting features that work across different data sources, then aligning those features globally and locally to improve classification accuracy while rejecting unknown data types.

This research paper describes a method for protecting privacy in distributed gradient descent (a technique where multiple computers work together to train AI models by each processing part of the data). The authors propose using hierarchical secret sharing (a cryptographic approach where information is split into pieces distributed across multiple parties, so no single party can see the complete data) to keep individual data private while still allowing the AI training process to work efficiently.

This research proposes a new approach to reinforcement learning (a machine learning technique where a system learns by trial and error) that guarantees robustness against uncertainties in how a system behaves and sensor measurements. The method uses neural networks (computational systems inspired by the brain) to model these uncertainties and identifies the weakest states during training so it can prioritize making them more reliable, with validation shown through quadrotor drone control tasks.

This research paper addresses a fundamental challenge in cooperative multiagent reinforcement learning (MARL, where multiple AI agents learn to work together toward shared goals). The authors propose UMARL, a new method that better balances exploration (trying new actions to learn more) and exploitation (using known good actions to earn rewards) by using local observations (information each agent can see from its own position). The method introduces specialized neural networks (agent representation network and individual weighting networks) to help agents learn optimal strategies even when they cannot see the entire environment.

This research paper, published in June 2026, presents a method for creating indexes in queryable-encrypted databases (databases where data stays encrypted even when being searched) that don't leak information about access patterns or query history. The approach aims to improve security by preventing attackers from inferring sensitive information about which data is being accessed based on observable patterns of database queries.