Majority Bit-Aware Watermarking for Large Language Models

Watermarks can help trace AI-generated text, but making them easier to detect can also make the writing less natural. MajorMark changes how identifying messages are woven into word choices so that more natural options remain available. The work aims to make text tracing more reliable while preserving readability, including when the embedded message is longer.

XMark: Reliable Multi-Bit Watermarking for LLM-Generated Texts

Identifying information can be hidden in AI-generated text to help trace where it came from. XMark makes that information easier to recover even from shorter passages, while limiting changes to the writing. The work supports practical attribution of generated content without requiring readers to notice a visible label or marker.

Traceable Black-Box Watermarks for Federated Learning

In collaborative AI training, participants receive copies of a shared model, creating a risk that a copy could be leaked. TraMark gives each participant's model a distinct watermark. This helps identify which copy was leaked by testing its responses, without inspecting its internal parameters, while aiming to preserve its usefulness for the original task.

Vision Token Manipulation Attacks on Cloud-Edge Inference of Large Vision-Language Models

Many AI systems analyze images on a device and send an intermediate representation to the cloud. This study shows that an attacker who alters part of that transmission can seriously disrupt the final answer. It identifies the communication link as a security weakness that must be considered when deploying image-and-language AI across devices and servers.

SelfGrader: LLM Jailbreak Detection via Anchored Token-Level Logits

AI assistants need to recognize attempts to bypass their safety rules without blocking ordinary questions. SelfGrader checks a model's internal scores for digits such as 0–9 to estimate how harmful a request is. This provides a lightweight safety check designed to reduce unnecessary refusals and avoid the variability of generating a separate written judgment.

On the Out-of-Distribution Backdoor Attack for Federated Learning

An attacker can teach a shared AI model to misread unusual but ordinary-looking inputs, without adding an obvious trigger. This study demonstrates that risk and introduces BNGuard, which checks for unusual internal patterns in participants' models. The goal is to detect poisoned contributions before they compromise a collaboratively trained model.

Detecting Backdoor Attacks in Federated Learning via Direction Alignment Inspection

Collaborative AI can be secretly sabotaged when a participant contributes an update that plants a hidden backdoor. AlignIns compares the overall direction and finer patterns of these updates to identify unusual behavior. Filtering suspicious contributions helps protect the shared model, including when participants have different kinds of training data.

Achieving Byzantine-Resilient Federated Learning via Layer-Adaptive Sparsified Model Aggregation

When many organizations or devices train an AI model together, a malicious participant can send an update intended to derail learning. LASA examines the important parts of each update layer by layer and filters suspicious contributions. This helps collaborative training remain reliable even when legitimate participants have very different datasets.

Identify Backdoored Model in Federated Learning via Individual Unlearning

A poisoned AI model may behave normally until a hidden trigger appears. MASA looks for these hidden backdoors by asking each participant's model to temporarily unlearn ordinary examples and examining how it reacts. The method helps distinguish malicious contributions from legitimate differences between participants before their updates enter the shared model.

Targeted Poisoning Attacks on Social Recommender Systems

Online recommendations can be manipulated through more than fake ratings. This study shows how attackers can combine fabricated ratings with social connections to steer recommendations toward particular users, including people they are not directly connected to. By exposing this weakness, the work helps clarify what future recommender systems need to defend against.

Heterogeneous Federated Fine-Tuning with Parallel One-Rank Adaptation

Devices with different computing budgets often struggle to improve a language model together. Fed-PLoRA breaks adaptation into small parallel modules, letting each device train what it can afford while the server combines their contributions more consistently. This makes collaborative language-model customization more accurate and efficient without collecting everyone's raw data in one place.

Fed-HeLLo: Efficient Federated Foundation Model Fine-Tuning With Heterogeneous LoRA Allocation

A laptop and a powerful server have different resources for improving a shared AI model. Fed-HeLLo assigns different parts of the model's lightweight adapters to different devices, based on available resources and which layers matter most. The result is more practical collaborative fine-tuning across unequal devices.

Fed-pilot: Optimizing LoRA Allocation for Efficient Federated Fine-Tuning with Heterogeneous Clients

Limited GPU memory can prevent smaller devices from joining collaborative AI training. Fed-pilot chooses which lightweight model components each device should train within its memory budget, then corrects imbalances when combining updates. This helps more devices contribute effectively while improving the balance between memory use and model quality.

Energy-Efficient Distributed Machine Learning at Wireless Edge with Device-to-Device Communication

When a device is short of energy, it can pass training data to a nearby device or wait until energy is available. This work coordinates those choices automatically, helping distributed devices learn a shared model with lower energy costs while maintaining learning performance.

Optimal Client Sampling in Federated Learning With Client-Level Heterogeneous Differential Privacy

People and organizations do not all need the same level of privacy protection. This work groups participants by their privacy needs and adjusts how often each group contributes to training. The shared AI model can learn more accurately while honoring each group's privacy requirements, without assuming the coordinating server can be fully trusted.

Trading Data For Learning: Incentive Mechanism For On-Device Federated Learning

Federated learning depends on people being willing to participate. This work links payments to participants' chosen privacy levels, so users can weigh compensation against privacy loss. It helps a model owner recruit useful contributors while respecting different privacy preferences, making collaborative learning more practical without requiring users to hand over their raw data.

Personalized Federated Learning With Differential Privacy

A single shared AI model may not work equally well for people with different habits or needs. This work lets devices learn personalized models while benefiting from one another's data without exposing the raw records. It adds formal privacy protection and accommodates differences in devices and participation, supporting more useful personalized services.

Federated Learning With Sparsified Model Perturbation: Improving Accuracy Under Client-Level Differential Privacy

Adding privacy protection can make a shared AI model less accurate, especially when it has many parameters. Fed-SMP reduces the amount of model information devices share before adding protective noise. This improves the balance between privacy and accuracy while lowering communication costs, making private collaborative learning more practical for connected devices.

Byzantine-Robust Federated Learning with Variance Reduction and Differential Privacy

Collaborative AI training must protect people's data while coping with participants who send harmful updates. FedVRDP tackles both challenges together: it limits what shared updates can reveal and reduces the disruption caused by malicious participants. Experiments show that a model can become more resistant to these attacks without giving up its formal privacy guarantee.

Federated Learning with Sparsification-Amplified Privacy and Adaptive Optimization

Devices can learn together without sharing raw data, but their updates may still reveal sensitive information. Fed-SPA shares only a randomly selected portion of each update and combines this with protective noise and faster learning. The result is a better balance between privacy, prediction accuracy, and the amount of information devices need to transmit.

Concentrated Differentially Private Federated Learning With Performance Analysis

Keeping data on a device does not automatically stop a shared AI model from revealing it. This work combines protective noise with secure aggregation, so the server sees combined updates instead of individual contributions. It also tracks privacy loss across repeated training, helping balance useful learning with measurable protection even when the server is not fully trusted.