Internet of Agents (IoA) envisions a unified, agent-centric paradigm where heterogeneous large language model (LLM) agents can interconnect and collaborate at scale. Within this paradigm, federated fine-tuning (FFT) serves as a key enabler that allows distributed LLM agents to co-train an intelligent global LLM without centralizing local datasets. However, the FFT-enabled IoA systems remain vulnerable to model poisoning attacks, where adversaries can upload malicious updates to the server to degrade the performance of the aggregated global LLM. This paper proposes a graph representation-based model poisoning (GRMP) attack, which exploits overheard benign updates to construct a feature correlation graph and employs a variational graph autoencoder to capture structural dependencies and generate malicious updates. A novel attack algorithm is developed based on augmented Lagrangian and subgradient descent methods to optimize malicious updates that preserve benign-like statistics while embedding adversarial objectives. Experimental results show that the proposed GRMP attack can substantially decrease accuracy across different LLM models while remaining statistically consistent with benign updates, thereby evading detection by existing defense mechanisms and underscoring a severe threat to the ambitious IoA paradigm.

Backdoor attacks pose a significant threat to the integrity of text classification models used in natural language processing. While several dirty-label attacks that achieve high attack success rates (ASR) have been proposed, clean-label attacks are inherently more difficult. In this paper, we propose three sample selection strategies to improve attack effectiveness in clean-label scenarios: Minimum, Above50, and Below50. Our strategies identify those samples which the model predicts incorrectly or with low confidence, and by injecting backdoor triggers into such samples, we aim to induce a stronger association between the trigger patterns and the attacker-desired target label. We apply our methods to clean-label variants of four canonical backdoor attacks (InsertSent, WordInj, StyleBkd, SynBkd) and evaluate them on three datasets (IMDB, SST2, HateSpeech) and four model types (LSTM, BERT, DistilBERT, RoBERTa). Results show that the proposed strategies, particularly the Minimum strategy, significantly improve the ASR over random sample selection with little or no degradation in the model's clean accuracy. Furthermore, clean-label attacks enhanced by our strategies outperform BITE, a state of the art clean-label attack method, in many configurations.

Poisoning attacks, in which an attacker adversarially manipulates the training dataset of a machine learning (ML) model, pose a significant threat to ML security. Beta Poisoning is a recently proposed poisoning attack that disrupts model accuracy by making the training dataset linearly nonseparable. In this paper, we propose four defense strategies against Beta Poisoning attacks: kNN Proximity-Based Defense (KPB), Neighborhood Class Comparison (NCC), Clustering-Based Defense (CBD), and Mean Distance Threshold (MDT). The defenses are based on our observations regarding the characteristics of poisoning samples generated by Beta Poisoning, e.g., poisoning samples have close proximity to one another, and they are centered near the mean of the target class. Experimental evaluations using MNIST and CIFAR-10 datasets demonstrate that KPB and MDT can achieve perfect accuracy and F1 scores, while CBD and NCC also provide strong defensive capabilities. Furthermore, by analyzing performance across varying parameters, we offer practical insights regarding defenses' behaviors under varying conditions.

Small language models (SLMs) are increasingly valued for their efficiency and deployability in resource-constrained environments, making them useful for on-device, privacy-sensitive, and edge computing applications. On the other hand, membership inference attacks (MIAs), which aim to determine whether a given sample was used in a model's training, are an important threat with serious privacy and intellectual property implications. In this paper, we study MIAs on SLMs. Although MIAs were shown to be effective on large language models (LLMs), they are relatively less studied on emerging SLMs, and furthermore, their effectiveness decreases as models get smaller. Motivated by this finding, we propose a new MIA called win-k, which builds on top of a state-of-the-art attack (min-k). We experimentally evaluate win-k by comparing it with five existing MIAs using three datasets and eight SLMs. Results show that win-k outperforms existing MIAs in terms of AUROC, TPR @ 1% FPR, and FPR @ 99% TPR metrics, especially on smaller models.