The latest A2A systems often use LLM-based agents for tool calling and planning workflows, introducing risks such as prompt injection attacks. We introduce a lightweight, proof-carrying receipt for agent-to-agent (A2A) messaging that supports fully offline verification. This receipt enables third parties to verify who authored a message and confirm whether it was logged, without contacting the sender, the receiver, or any online service. An agent signs a canonical JSON representation of each message. Servers log the payload hash in a Merkle tree and periodically sign the tree state as a Signed Tree Head (STH). The returned bundle includes a receipt containing a Merkle proof, an STH, and a minimal decentralized identifier (DID) registry for offline key discovery. A verifier recomputes the Merkle root and validates the payload-hash match, Merkle inclusion, and both the STH and client signatures entirely offline. In our prototype, the proof size grows as $O(\log N)$, and the median offline verification time is approximately 0.4 s on our test platform. Receipts self-authenticate individual messages within a single transparency log, providing a lightweight alternative to other approaches.
Traditional data anonymization techniques-including generalization, suppression, perturbation, and masking-were designed for an era of limited contextual modeling. These methods effectively balanced privacy protection and data utility for structured, textual, audio, and visual data. However, the emergence of Large Language Models (LLMs) and multimodal foundation architectures has fundamentally disrupted this balance. Modern models, equipped with powerful representation-learning and cross-modal reasoning, can infer or reconstruct masked identities using contextual, semantic, or visual cues, rendering traditional anonymization insufficient. This paper surveys anonymization techniques across tabular, textual, audio, and video domains, comparing their effectiveness before and after the advent of LLMs. Through controlled experiments, we demonstrate that context-aware LLMs can accurately re-identify subjects from anonymized datasets, even when explicit identifiers are removed. The results reveal that traditional anonymization fails to address contextual inference threats, calling for a paradigm shift toward holistic, multi-layered, privacy-preserving approaches. We discuss emerging strategies-such as differential privacy, federated learning, synthetic data generation, and LLM-in-the-loop anonymization-as potential solutions for privacy preservation in the LLM era.
Given the ubiquity of computers in nearly all work environments, including highly secure facilities and classified systems, it is essential to accurately identify the current user operating a computer to ensure the security of both data and control systems associated with workstations. This paper investigates the efficacy of various machine learning algorithms in distinguishing users based on mouse and keyboard dynamics. Four algorithms are employed for user classification: K-Nearest Neighbor, Decision Tree, Artificial Neural Networks, and Convolutional Neural Networks. The study utilizes two primary datasets, one for mouse dynamics and one for keyboard dynamics. These datasets are pre-processed using Principal Component Analysis (PCA), then divided into training and test subsets before applying the machine learning algorithms. The methods outlined herein demonstrate promise in user identification through analysis of interactions with a computer's mouse and keyboard.
This paper introduces DP-Prox, an end-to-end framework for robust and differentially private federated instruction tuning of small Large Language Models (LLMs) on 8 GB edge devices. DP-Prox is instantiated using the 3.8 billionparameter Phi-3-mini model with 4-bit Quantized Low-Rank Adaptation (QLoRA), and combines (i) FedProx-style proximal regularization on low-rank adapters to stabilize optimization under highly Non-IID smart-home data, and (ii) Differentially Private SGD (DP-SGD) applied directly to adapter gradients to provide user-level $(\epsilon, \delta)$-differential privacy against an honest-but-curious server. In a realistic IoT instruction-following benchmark with Dirichlet-skewed clients ($\alpha=0.3$), DP-Prox improves F1-score from 0.80 (Edge-FIT with FedAvg) to 0.82 while reducing convergence rounds from $T=50$ to $T \approx 40$. Under a strict privacy budget $\left(\epsilon=5, \delta=10^{-5}\right)$, DP-Prox maintains an F1-score of 0.77 with only $+15 \%-20 \%$ local overhead, and preserves the communication efficiency of QLoRA-based federated training. Qualitative analyses further show that DP-Prox yields more consistent instruction-following behavior across heterogeneous homes, especially on multi-device routines and privacy-sensitive queries. Together, these results position DP-Prox as a practical and regulation-ready blueprint for deploying small, edge-resident LLMs in smart-home IoT environments.
We study post-clue-finding coordination for bandwidth- and compute-limited multi-robot search. Following standard Bayesian search theory, each grid cell is treated as a segment with an associated prior probability of area (POA), and detection of a clue triggers a posterior update using a Manhattan-distance decay centered on the clue cell. We introduce a single-message “can-win” task-cell-reservation rule and implement it on two low-cost Pololu 3pi+ 2040 robots communicating via MQTT. The proposed rule designates a robot as eligible to reserve a task-cell when its distance-adjusted posterior POA value exceeds that of all peers for the same candidate cell, enabling single-message coordination without multi-round bidding and approximating POS-maximizing effort allocation under strict bandwidth limits. We compare two algorithms: (A) reservation with a distance-aware value function and (B) a probability-only greedy baseline without reservation, with identical planning and costs otherwise. Real-world and matched simulation experiments show consistent improvements with reservation in post-clue coverage efficiency (up to +4.7 percentage points), reduced revisits, and lower run-to-run variability, with only a 6 % increase in per-step message traffic. Simulation reproduces IRL totals within $\pm 3$ steps (TOST, $p<0.03$) and exhibits high concordance ($\mathbf{C C C}>0.97$). Results suggest that a minimal reservation broadcast can stabilize decentralized Bayesian search coordination for embedded search platforms operating under tight communication constraints.
