Out-of-distribution (OOD) detection is crucial to safety-critical machine learning applications and has been extensively studied. While recent studies have predominantly focused on classifier-based methods, research on deep generative model (DGM)-based methods have lagged relatively. This disparity may be attributed to a perplexing phenomenon: DGMs often assign higher likelihoods to unknown OOD inputs than to their known training data. This paper focuses on explaining the underlying mechanism of this phenomenon. We propose a hypothesis that less complex images concentrate in high-density regions in the latent space, resulting in a higher likelihood assignment in the Normalizing Flow (NF). We experimentally demonstrate its validity for five NF architectures, concluding that their likelihood is untrustworthy. Additionally, we show that this problem can be alleviated by treating image complexity as an independent variable. Finally, we provide evidence of the potential applicability of our hypothesis in another DGM, PixelCNN++.

The training of score-based diffusion models (SDMs) is based on score matching. The challenge of score matching is that it includes a computationally expensive Jacobian trace. While several methods have been proposed to avoid this computation, each has drawbacks, such as instability during training and approximating the learning as learning a denoising vector field rather than a true score. We propose a novel score matching variant, local curvature smoothing with Stein's identity (LCSS). The LCSS bypasses the Jacobian trace by applying Stein's identity, enabling regularization effectiveness and efficient computation. We show that LCSS surpasses existing methods in sample generation performance and matches the performance of denoising score matching, widely adopted by most SDMs, in evaluations such as FID, Inception score, and bits per dimension. Furthermore, we show that LCSS enables realistic image generation even at a high resolution of $1024 \times 1024$.

The task of out-of-distribution (OOD) detection is vital to realize safe and reliable operation for real-world applications. After the failure of likelihood-based detection in high dimensions had been shown, approaches based on the typical set have been attracting attention; however, they still have not achieved satisfactory performance. Beginning by presenting the failure case of the typicality-based approach, we propose a new reconstruction error-based approach that employs normalizing flow (NF). We further introduce a typicality-based penalty, and by incorporating it into the reconstruction error in NF, we propose a new OOD detection method, penalized reconstruction error (PRE). Because the PRE detects test inputs that lie off the in-distribution manifold, it effectively detects adversarial examples as well as OOD examples. We show the effectiveness of our method through the evaluation using natural image datasets, CIFAR-10, TinyImageNet, and ILSVRC2012.

Fully homomorphic encryption (FHE) is a promising tool for privacy-preserving applications, and it enables us to perform homomorphic addition and multiplication on FHE ciphertexts without decrypting them. FHE has two types: one supporting the exact computation and the other supporting the approximate computation. Further the FHE schemes supporting the exact computation have two types, bit-wise FHE, which encrypts a plaintext bit by bit, and integer-wise FHE, which encrypts a plaintext as an integer. Both types of FHE are important depending on the types of computation we need to execute securely. In this work, we focus on integer-wise FHE, and propose improved methods for integer-wise homomorphic comparison and division operations. For a comparison operation, we propose a method that halves the number of necessary homomorphic multiplications by introducing an odd function as an interpolated polynomial to be evaluated, as opposed to the previous work of Narumanchi et al. (AINA ’17). For a division operation, as opposed to the previous work of Okada et al. (WISTP ’18), we propose a simple method to reduce the processing time by introducing an equality function based on Fermat’s little theorem without changing the multiplicative depth, and show the analysis of why this approach can achieve better efficiency in detail. In our homomorphic division, the number of interpolated polynomials is reduced by half, thus also achieving the reduction of the processing time of precomputations and the number of polynomials to be stored. We also implement our improved methods in HElib, which is one of popular FHE libraries using the BGV encryption. As a result, we show that, e.g., in the plaintext space , our homomorphic comparison with the Paterson-Stockmeyer method is faster by a factor of about 5.61 compared with Narumanchi et al. (AINA ’17) and our homomorphic division is faster by a factor of about 1.45 compared with Okada et al. (WISTP ’18).

Abstract Searchable encryption allows us to perform a keyword search over encrypted data. However, we cannot efficiently perform some complex search (e.g., a wildcard search) with traditional searchable encryption schemes because they can deal with only equality matches. Our symmetric searchable encryption can deal with partial matches. This allows us to efficiently perform a wildcard search, partial match search, and so on. We also examine the feasibility of our scheme by experiments on a smartphone and tablet, and confirm our scheme can be used in these environments. Availability on portable devices will offer high convenience. Copyright © 2013 John Wiley & Sons, Ltd.

Password-based Authenticated Key Exchange (PAKE) allows a server to authenticate a user and to establish a session key shared between the server and the user just by having memorable passwords. In PAKE, conventionally the server is assumed to have the authentication functionality and also provide on-line services simultaneously. However, in the real-life applications, this may not be the case, and the authentication server may be separate from on-line service providers. In such a case, there is a problem that a malicious service provider with no authentication functionality may be able to guess the passwords by interacting with other participants repeatedly. Abdalla et al. put forward a notion of the server password protection security to deal with this problem. However, their proposed schemes turned out to be vulnerable to Undetectable On-line Dictionary Attack (UDonDA). To cope with this situation, we propose the Gateway Threshold PAKE provably secure against this password guessing attack by also taking the corruption of authentication servers into consideration.

In online services, e.g., Online shopping, a service provider (SP) manages access logs containing customers' buying histories. Therefore, user's personal information, e.g., Their hobbies and diversions, is revealed from the exposed logs if each customer can be linked. In fact, such information exposure has occurred due to the popularization of online services. To cope with this problem, SPs may only have to delete access logs, but then no illegitimate users, who accessed the server illegally, will be traced from the logs. In this paper, we propose a log management mechanism where (1) no user information is revealed even if logs are exposed, but (2) illegitimate users can be traced when necessary. Specifically, we consider single sign on (SSO) systems, since plural access logs might be connected by one account, and this could trigger the above privacy infringement problem. We construct our privacy-enhanced access log management mechanism based on the Wang-Wang-Susilo SSO system (TrustCom 2013) which applies nominative signatures as its building block. Specifically, we realize the system by additionally applying the invisibility property of the Schuldt-Hanaoka nominative signature scheme (ACNS 2011). Finally, we estimate the efficiency of the proposed system by using Pairing-Based Cryptography (PBC) library and confirmed that for each algorithm, computation time is at most just over 80 milliseconds on a PC, which seems sufficiently practical.

In privacy‐carrying email systems, the authors should guarantee that the email content is confidential and sometimes the sender/receiver identities are hidden. Also, they require that the key generation is flexible and manageable. In this study, first, they propose an anonymous encryption scheme that supports a partial‐order subset delegatable ability. The proposed scheme achieves the security properties of confidentiality against adaptive chosen‐plaintext attacks, anonymity against adaptive chosen‐subset attacks and computational delegation indistinguishability. Secondly, they provide a deployment application of their anonymous encryption in an interdisciplinary group email management system with flexible and fine‐grained key delegation. The deployment can achieve the privacy of message confidentiality, receiver anonymity and delegation obliviousness, which has fine‐grained security in secure email systems. Finally, they provide an extension for the chosen‐ciphertext secure scheme, and discuss the efficiency for decryption and the security level.