Yen-Lung Lai

YenLung Lai, Xingbo Dong, Zhe Jin

In this research, we introduce the concept of "computational entanglement," a phenomenon observed in overparameterized feedforward linear networks that enables the network to achieve zero loss by fitting random noise, even on previously unseen test samples. Analyzing this behavior through spacetime diagrams reveals its connection to length contraction, where both training and test samples converge toward a shared normalized point within a flat Riemannian manifold. Moreover, we present a novel application of computational entanglement in transforming a worst-case adversarial examples-inputs that are highly non-robust and uninterpretable to human observers-into outputs that are both recognizable and robust. This provides new insights into the behavior of non-robust features in adversarial example generation, underscoring the critical role of computational entanglement in enhancing model robustness and advancing our understanding of neural networks in adversarial contexts.

Yen-Lung Lai, Zhe Jin

This work investigates the generalization behavior of deep neural networks (DNNs), focusing on the phenomenon of "fooling examples," where DNNs confidently classify inputs that appear random or unstructured to humans. To explore this phenomenon, we introduce an analytical framework based on maximum likelihood estimation, without adhering to conventional numerical approaches that rely on gradient-based optimization and explicit labels. Our analysis reveals that DNNs operating in an overparameterized regime exhibit a collapse in the output feature space. While this collapse improves network generalization, adding more layers eventually leads to a state of degeneracy, where the model learns trivial solutions by mapping distinct inputs to the same output, resulting in zero loss. Further investigation demonstrates that this degeneracy can be bypassed using our newly derived "wormhole" solution. The wormhole solution, when applied to arbitrary fooling examples, reconciles meaningful labels with random ones and provides a novel perspective on shortcut learning. These findings offer deeper insights into DNN generalization and highlight directions for future research on learning dynamics in unsupervised settings to bridge the gap between theory and practice.

Yen-Lung Lai

Secure sketch produces public information of its input $w$ without revealing it, yet, allows the exact recovery of $w$ given another value $w'$ that is close to $w$. Therefore, it can be used to reliably reproduce any error-prone secret (i.e., biometrics) stored in secret storage. However, some sources have lower entropy compared to the error itself, formally called "more error than entropy", a standard secure sketch cannot show its security promise perfectly to these kind of sources. This paper focuses on secure sketch. We propose a concrete construction for secure sketch. We show security to all noisy sources, including the trivial source with zero min-entropy. In addition, our construction comes with efficient recovery algorithm operates in polynomial time in the sketch size, which can tolerate high number of error rate arbitrary close to 1/2. Above result acts in conjunction to our derivation on the solution to two NP-complete coding problems, suggesting P=NP.

Yen-Lung Lai, Jung-Yeon Hwang, Zhe Jin et al.

In this paper, we propose a novel biometric cryptosystem for vectorial biometrics named symmetric keyring encryption (SKE) inspired by Rivest's keyring model (2016). Unlike conventional biometric secret-binding primitives, such as fuzzy commitment and fuzzy vault, the proposed scheme reframes the biometric secret-binding problem as a fuzzy symmetric encryption problem with a notion called resilient vector pair. In this study, the pair resembles the encryption-decryption key pair in symmetric key cryptosystems. This notion is realized using the index of maximum hashed vectors - a special instance of the ranking-based locality-sensitive hashing function. With a simple filtering mechanism and [m,k] Shamir's secret-sharing scheme, we show that SKE, both in theoretical and empirical evaluation, can retrieve the exact secret with overwhelming probability for a genuine input yet negligible probability for an imposter input. Though SKE can be applied to any vectorial biometrics, we adopt the fingerprint vector as a case of study in this work. The experiments have been performed under several subsets of FVC 2002, 2004, and 2006 datasets. We formalize and analyze the threat model of SKE that encloses several major security attacks.

Zhe Jin, Yen-Lung Lai, Andrew Beng Jin Teoh

Despite a variety of theoretical-sound techniques have been proposed for biometric template protection, there is rarely practical solution that guarantees non-invertibility, cancellability, non-linkability and performance simultaneously. In this paper, a ranking-based representational transformation is proposed for fingerprint templates. The proposed method transforms a real-valued feature vector into index code such that the pairwise-order measure in the resultant codes are closely correlated with rank similarity measure. Such a ranking based technique offers two major merits: 1) Resilient to noises/perturbations in numeric values; and 2) Highly nonlinear embedding based on partial order statistics. The former takes care of the accuracy performance mitigating numeric noises/perturbations while the latter offers strong non-invertible transformation via nonlinear feature embedding from Euclidean to Rank space that leads to toughness in inversion. The experimental results demonstrate reasonable accuracy performance on benchmark FVC2002 and FVC2004 fingerprint databases, thus confirm the proposition of the rank correlation. Moreover, the security and privacy analysis justify the strong capability against the existing major privacy attacks.