In briefCalibrating Difficulty Over Time for Diffusion Model Membership Inference
SummaryThis paper studies membership inference attacks, which test whether an image was used to train a diffusion model. It argues that raw reconstruction loss is unreliable because some images are inherently easier to reconstruct, and proposes calibrating for this difficulty. Reference-Model Calibration compares a fine-tuned model against an earlier checkpoint, while Temporal Self-Calibration uses losses at neighboring diffusion timesteps from a single model. The authors report that both methods outperform prior attacks and generalize across x-, v-, and ε-prediction models.
In briefTextSeal2: Distortion-Free LLM Watermarking with Copula Mixing
SummaryThe paper presents TextSeal2, a distortion-free watermark for large language models that keeps its signal under speculative decoding without changing the acceptance rate. It preserves output diversity and adds no measurable inference latency. The method merges multiple keys into one virtual key coupled to fresh noise via a Gaussian copula. Evaluations on three LLMs report higher detection rates than its predecessor TextSeal at equal output diversity.
In briefPower-adjusting attacks on mining pools under block-dependent payout schemes
SummaryThis paper analyzes the T-PAW power-adjusting block withholding attack under three payout schemes used in practice: PPLNS, round-based PROP, and exponential-decay EXP. It also introduces Power Adjusting Hopping (PAH), in which an attacker adjusts its power toward the pool according to round blockwork age without withholding blocks. The authors derive revenue ratios analytically and verify them with Monte Carlo simulations. They find that PROP is highly vulnerable to PAH, while PPLNS is the least vulnerable scheme.
WhyThe authors analyze power adjusting attacks under block-dependent payout schemes, propose PAH, derive revenue ratios analytically, and report Monte Carlo verification.
In briefPrivacy and security risks when optimizing large language models for edge devices
SummaryThis paper examines optimization techniques used to deploy large language models on edge devices, aimed at reducing memory, computation, and energy costs. The authors argue that preserving task performance does not necessarily preserve security. They contend that optimization transforms information in ways that can alter representations, decision pathways, and trust boundaries, creating security drift and new attack surfaces introduced by the optimization itself.
Mingchen Li, Ren-Yi Huang, Dumindu Samaraweera et al.
In briefEfficient oversight of AI computations through rewinding
SummaryThis paper studies how a weak verifier can check computations performed by more powerful AI systems. It introduces rewindable interactive proofs (RIPs), in which the verifier can rewind the prover to an earlier state and continue from there. The authors build a relativizing RIP for computations whose outputs are robust to changes in a small fraction of oracle answers, using no cryptography. The construction is more efficient than prior methods that do not assume a truthful prover, and a preliminary MNIST experiment is reported.
In briefRecHash: Segmented Hash-Entangled Chains for Stateful Sequential Authentication
SummaryThis paper presents RecHash, a stateful sequential authentication scheme built on paired hash-chain openings and an authenticated record history. It extends the 2019 hash-entangled construction by assigning each record to a predetermined chain segment, so every digest block fits its scheduled slot. The paper gives a segmented state transition, a worked example, allocation and resource results, and a renewal-free security composition theorem under authenticated history and an adaptive one-step guarantee.
In briefDeep learning models for continuous typing-based identity checks
SummaryThis paper proposes a two-level keystroke dynamics authentication system. The first level uses Temporal Convolutional Networks with contrastive learning and population-based training to verify fixed-text passwords, reaching a 3.77% equal error rate. After login, the second level monitors typing during free-text sessions with a four-model deep learning ensemble. If suspicious behavioral drift is detected, the user is sent back to the first level for re-authentication. The free-text component reports a 10.05% equal error rate.
Suprava Saha Dibya, Minar, Abdulla Al Noman et al.
In briefUnderstanding and Mitigating Role Spoofing in LLMs
SummaryThis paper studies how large language models track who said what in a conversation, using both content cues such as style and structural cues such as control tokens. It shows that models are highly vulnerable to role spoofing attacks that combine fake structure and content. In a toy setting, adversarial training against content-based injections mitigates those attacks but leaves models susceptible to structure-based ones. The authors also propose a detection method using linear probes trained to predict conversational roles, which they report is effective across different settings.
Ezra Edelman, Benjamin Edelman, Surbhi Goel et al.
