Game Theory and LLMs Team Up to Waste Phone Scammers' Time
A new arXiv preprint proposes using LLM-driven 'bot wars' to impose real costs on phone scammers.
Research Digest
Preprints and papers, rewritten in plain English.
Secure Aggregation Limits in Federated Learning with Dropouts
A new theoretical study maps the fundamental trade-offs in secure aggregation when users drop out during federated learning.
Kernel-Level Escrow Delays AI Agent Egress Until a Verifier Rules
A proposed mechanism holds container network egress in the kernel until an isolated verifier approves, closing bypasses used by child processes and direct sockets.
Study examines how LLM calls disrupt software contracts
LLM calls in agentic systems challenge traditional software quality.
End-of-Life Care Expands Digital Data While Patient Control Declines
A new arXiv study examines how terminally ill patients generate extensive digital health data but face shrinking control over it.
Hidden Integer Feedback: Privacy Loss Across Successive Releases
A new arXiv paper examines how hidden randomized rounding couples privacy loss in sequential integer releases, proposing a rational-feedback mechanism.
Exact GPU Correlation for Auditing Firewall Logs
A new preprint proposes a GPU-based method for auditing firewall logs against compliance requirements while preserving uncertainty from missing attributes.
Q-MALDetect: Quantum Image-Based Malware Detection and Adversarial Robustness
A new paper introduces Q-MALDetect, a hybrid quantum-classical model for image-based malware detection, and probes its adversarial robustness.
Closing the Gap in Secure Compilation Proofs for Compartmentalized C
A new approach aims to make formally verified compilers for compartmentalized low-level code account for real protection mechanisms and memory models.
Topic-Based Watermarking for Peer Reviews: A Feasibility Study
A new arXiv paper examines whether topic-based watermarking can help identify AI-generated peer reviews.
Federated Learning With Differential Privacy: Balancing Privacy and Utility
A new method targets the trade-off between privacy protection and model performance in federated learning for heterogeneous clients.
BitPCE: GPU Bitmap Algebra for Multi-Dimensional Policy Resolution
A new arXiv preprint proposes using batched GPU bitmap operations to speed up policy controllers that match workloads against multi-dimensional rules.
Brush stroke re-rendering defeats invisible watermarks on digital paintings
A simple stroke-based redrawing attack can strip invisible ownership watermarks from digital paintings, according to a new preprint.
Hand Drawing as a Tool for Framing Privacy and Security Research
A new preprint reports on a classroom exercise in which students used collaborative hand drawing to think through usable privacy and security research problems.
Revisiting Secret Sharing for When Encryption Keys Leak
A new arXiv paper re-examines computational secret sharing as a defense against data exposure when encryption keys fall into an attacker's hands.
ZeroDSS Segments Images Using Only Class Names
A new method performs semantic segmentation without any real images or pixel-level labels, relying solely on class names.
Offline CBDC Design Principles Laid Out in New Paper
A recent arXiv paper identifies three major objectives and corresponding countermeasures for central bank digital currencies that need to function offline.
Watermarking LLM Agents' Decision Sequences for Provenance
A new approach embeds identifiable markers in the sequential decisions of LLM agents, making it easier to trace provenance and deter unauthorized reuse.