Research

Researchers present ShadowCoT, a backdoor attack framework that targets the internal chain-of-thought reasoning of LLMs rather than tokens or prompts. By conditioning on internal reasoning states and rewiring attention pathways, updating only 0.15% of parameters, it hijacks multi-step reasoning and produces stealthy adversarial CoTs. Across reasoning benchmarks and LLMs it reports an average Attack Success Rate of 91.2% (peak 94.4%) and a Hijacking Success Rate of 84.9%.

Researchers propose a secure off-chain payment protocol (SOCP) for blockchain payment-channel networks that addresses interception attacks on the Lightning Network, where hash preimages travel in plaintext and malicious intermediate nodes can claim funds without passing them on. The protocol is built on a new primitive, the ciphertext unlinkable autonomous path proxy re-encryption scheme (CUAP-PRE), which lets the delegator designate only trusted delegatees and resists homology inference and delegation path tracing. The authors claim post-compromise security under the DBDH assumption in the random oracle model, supported by formal proofs and evaluations.

Researchers present a method for statistically validating physical prime random number generators, which existing test suites cannot assess because they target uniform pseudorandom sources. They extend an equiprobable binning framework of the prime distribution with a predictive machine learning model that learns correlations in prime output. The framework identifies flawed configurations of a prime QRNG based on vacuum-state quadrature measurements for output lengths up to 128 bits, and shows that a classical electronic-noise source produces a conclusively biased prime distribution, while the QRNG remains resilient to this cryptanalysis.

This paper shows that existing internally sampled honeyword schemes, Honeyindex (TDSC'16) and Superword (COSE'21), contain two critical security flaws and two critical usability flaws. Using Bayesian and intersection attack theories against three attacker models, the authors find that with 40 sweetwords per user, a basic attacker identifies 4.31%~5.04% of real passwords in 10^4 honeyword login attempts, above the ideal 2.50% (1/40), and more advanced attackers identify 18.6%~44.8% and 20.6%~43.6%.

This paper proposes MCCENet, a deep multimodal fusion network for palm authentication that combines palm vein (PV) and palm shape (PS) data. The authors observe that discriminative palm shape contour information is mainly captured in shallow layers, so they design a hierarchical feature fusion (HFF) module for inter-modal channel exchange at those layers, plus a multimodal contrastive learning loss to align features across modalities. Experiments on eight public datasets report state-of-the-art performance in all cases.

FDXT is a new conjunctive dynamic searchable symmetric encryption scheme that guarantees forward privacy for all queried keywords in a multi-keyword search, whereas ODXT only protects the single first queried keyword. It also avoids the volume leakage that SDSSE-CQ suffers from, and experiments on the Crime, Wikipedia and Enron datasets show performance comparable to ODXT and significantly better than SDSSE-CQ.

The paper introduces TMVS, a software-only Threshold-based Majority Voting Scheme that reduces noise and bias in SRAM PUF key generation without cell-level bit-error qualification or SRAM redesign. The authors derive closed-form decoding-error and memory expressions, validate them on experimental data, and show near-zero cross-chip secrecy leakage with average conditional min-entropy above 1 bit. In a representative configuration, TMVS generates a 128-bit key with failure probability 9.15 x 10^-7 and zero secrecy leakage at a 10% bit-flip probability, using about 248k clock cycles on a 32-bit ARM Cortex-M0.

Researchers propose Monte Carlo marginalization (MCMarg), a differentiable method that fits a Gaussian mixture model to an intractable high-dimensional distribution by minimizing KL divergence, with kernel density estimation keeping the optimization differentiable. The method reports FID scores of 22 on MNIST without a neural network and 2.69 on CIFAR-10, and it improves pretrained VAEs by about 10 FID points when replacing their standard priors.

The paper proposes a Conceptual Framework for AI for Business (AI4B) Competency Areas to guide business school curricula. It lists seven competency areas, including AI Use and AI Literacy, AI Governance and Risk Management, and AI Systems Development. An initial analysis of AI-related programs at leading U.S. business schools provides early validation for the framework.