Transfer learning is a de facto standard method for efficiently training machine learning models for data-scarce problems by adding and fine-tuning new classification layers to a model pre-trained on large datasets. Although numerous previous studies proposed to use homomorphic encryption to resolve the data privacy issue in transfer learning in the machine learning as a service setting, most of them only focused on encrypted inference. In this study, we present HETAL, an efficient Homomorphic Encryption based Transfer Learning algorithm, that protects the client's privacy in training tasks by encrypting the client data using the CKKS homomorphic encryption scheme. HETAL is the first practical scheme that strictly provides encrypted training, adopting validation-based early stopping and achieving the accuracy of nonencrypted training. We propose an efficient encrypted matrix multiplication algorithm, which is 1.8 to 323 times faster than prior methods, and a highly precise softmax approximation algorithm with increased coverage. The experimental results for five well-known benchmark datasets show total training times of 567--3442 seconds, which is less than an hour.

We propose a novel computer architectural concept of instruction overloading to support block ciphers. Instead of adding new instructions, we extend only the execution of some existing instructions. The proposed method allows a central processing unit core to execute different operations for the same instructions, depending on the address of the data, similar to operator overloading in object-oriented languages. We first present an extension for the AES algorithm, then we demonstrate its enhanced applicability with two further extensions supporting multiple block ciphers and hardware masking. The first extension for AES is also applicable to add/AND-rotate-XOR-based block ciphers such as SIMON. The AES and SIMON encryption speed, on this extended core, is at least doubled and is significantly less affected by memory latency. In addition, the AES encryption code requires only 18% of the memory of the previous software implementation. The second extension can further support various block ciphers defined over GF(28), and the SM4 encryption speed is increased by at least 182%. The third extension provides correlation power analysis (CPA) resistance with a 66.6% area overhead but almost no speed overhead, whereas a typical software anti-CPA AES implementation requires at least hundreds of times the execution time.

In 2018, Zhou et al. proposed PassBio, a privacy-preserving biometric authentication scheme, in IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY (TIFS) (DOI: 10.1109/TIFS.2018.2838540). PassBio delivers its promise and effectively protects users’ privacy from malicious servers. However, we demonstrate that PassBio may be vulnerable in a more inclusive attack model with malicious clients.

Montgomery reduction is a well-known optimization technique for modular reduction over prime fields. However, it has rarely been used for national institute of standards and technology (NIST) prime fields, because the special structure of NIST primes enables fast reduction that requires fewer computations than Montgomery reduction. In this study, we rediscover the Montgomery reduction for NIST primes and propose a new modular reduction method by applying partial modular reduction to Montgomery reduction. Using this new modular reduction method and other optimization techniques, such as pipelining and parallel processing, we propose an efficient elliptic curve cryptography (ECC) coprocessor over NIST prime fields. The implementation results indicate that the proposed coprocessor can perform one elliptic curve point multiplication in 0.055 ms with only 194.7 k gate counts. This is a considerably higher speed per area compared to previous research.

We propose a privacy-preserving face authentication system for smartphones that guarantees security against malicious clients. Using the proposed system, a face feature vector is stored on a remote server in encrypted form. To guarantee security against an honest-but-curious server who may try to learn the private feature vector, we perform a Euclidean distance-based matching score computation on encrypted feature vectors using homomorphic encryption. To provide security against malicious clients, we adopt a blinding technique. We implement the proposed system on a mobile client and a desktop server. Through an experiment with real-world participants, we demonstrate that secure face verification can be completed in real time (within 1.3 s) even when a smartphone is involved, with an Equal Error Rate (EER) of 3.04%. In further experiments with two public face datasets, CFP and ORL, face verification is completed in approximately 1 s with EER of 1.17% and 0.37%, respectively. Our system is two orders of magnitude faster than previous privacy-preserving face verification method with the same security assumptions and functionalities. To achieve this secure real-time computation, we improve the Catalano-Fiore transformation which converts a linear homomorphic encryption scheme into a quadratic scheme, and parallelize the decryption procedure of our system.

A Personal Identification Number (PIN) is a multiple-digit sequence widely used for user authentication . It is desirable for a PIN-entry method to be secure against two main security threats, random guessing attacks and recording attacks. Although there have been many proposals for challenge-response-based PIN-entry methods, it is well known that the only way to prevent both attacks is to physically prevent attackers from observing the challenge-response pairs, which motivates the development of PIN-entry methods that use secure secondary channels such as audio signals . To provide a guideline for designing an audio-based PIN-entry method, we propose a simple framework to transform a non-audio-based method into an audio-based one. We also present a new PIN-entry method that improves the performance of this simple transformation. Most audio-based methods in the literature are unimodal methods, that is, they transmit almost all required information through an audio channel because it was believed that this approach maximized the user's performance. In this paper, however, we show that a carefully designed bimodal system may be more usable than a unimodal one. We present a new PIN-entry method whose audio channel only transmits the minimal required data while most of the challenge information is transmitted through the efficient visual channel. Our user study shows that the PIN-entry time of the proposed method is shorter than those of the previous audio-based methods, while its error rate is kept as low as that of the previous methods.

ABSTRACT Since Rathaet al. introduced the functional transform for cancelable fingerprint templates, a few simulation attacks to this method have been proposed. The attacks are based on the fact that the transform depends only on the predefined parameters. That is, the attacker may fully simulate the transform and partially invert it if the parameters are available. Although an original template is transformed using different parameters for different systems, even the compromise of only one of these systems may reveal the original template, which may be a serious potential threat from a practical viewpoint. In this paper, we propose an improved functional transform, whose parameters are decided by the original template, as well as predefined user‐specific keys. Because the information on the original template will not be available to the attacker even when a system is compromised, the proposed method significantly improves the security of the original template by preventing the attacker from reconstructing the transform. Copyright © 2013 John Wiley & Sons, Ltd.

The personal identification number (PIN) is a well-known authentication method used in various devices, such as ATMs, mobile devices, and electronic door locks. Unfortunately, the conventional PIN-entry method is vulnerable to shoulder-surfing attacks. Consequently, various shoulder-surfing resistant methods have been proposed. However, the security analyses used to justify these proposed methods are not based on rigorous quantitative analysis, but instead on the results of experiments involving a limited number of human attackers. In this paper, we propose new theoretical and experimental techniques for quantitative security analysis of PIN-entry methods. We first present new security notions and guidelines for secure PIN-entry methods by analyzing the existing methods under the new framework. On the basis of these guidelines, we develop a new PIN-entry method that effectively obviates human shoulder-surfing attacks by significantly increasing the amount of short-term memory required in an attack.

In 2008, Dimitrov et al. proposed a point multiplication algorithm on Koblitz curves using multiple-base expansions. They claimed that their algorithm is the first provably sublinear point multiplication algorithm on Koblitz curves. In this paper, we show that the well-known τ-adic NAF method is already sublinear and also guarantees a better average performance.

We propose a new method to speed up discrete logarithm (DL)-based cryptosystems by considering a new variant of the DL problem, where the exponents are formed as e1+ e2for some fixed a and two integers e1, ae2with a low weight representation. We call this class of exponents split exponents, and we show that with certain choice of parameters the DL problem on split exponents is essentially as secure as the standard DL problem, while the exponentiation operation using exponents of this class is significantly faster than best exponentiation algorithms given for standard exponents. For example, the speed of scalar multiplication on the standard Koblitz curve K163 is estimated to be accelerated by up to 51.5 % and 23.5 % at the cost of memory for one precomputed point, compared to the TNAF and window TNAF methods, respectively. As for security, we show that the provable security of the DL problem using split exponents is only by a small constant, e.g., 1/4, worse than the security of the standard DL problem. Split exponents can be adopted to speed up various DL-based cryptosystems. We exemplify this on the recent CCA-secure public key encryption of Bellare, Kohno, and Shoup.

Vonzur Gathen proposed an efficient parallel exponentiation algorithm in finite fields using normal basis representations. In this paper we present a processor-efficient parallel exponentiation algorithm in GF(2 n ) which improves upon von zur Gathen's algorithm. We also show that exponentiation in GF(2 n ) can be done in Ο(log n) time using n/(log n)2 processors. Hence we get processor x time bound Ο(n/log n), which is optimal. Finally, we present an on-line processor assignment scheme which was missing in von zur Gathen's algorithm, and show that its time complexity is negligible.