In our Secure Computing (SC) group, we conduct research to process data with preserving privacy and secret information in the data. We focus on two technologies that are secure computation and differential privacy.
For example, consider a scenario where a company wants to analyze their data to extract information, but the company does not have a technology or computational resource to do so. In such scenario, the company can outsource such computation to a third-party such as cloud. However, if the data contains important information for the company or personal information, it is insecure to simply outsource the data and computation to the third-party as they may be malicious or under attack. Our research goal is to avoid this risk.
One of secure computation technology is homomorphic encryption. A client encrypts their data with homomorphic encryption and outsources it to a third-party for analysis. Next, the third-party processes the encrypted data without decryption and send an encrypted result to the client. Finally, only the client can decrypt the encrypted result. In these processes, no secret information is revealed. However, homomorphic encryption has some problems such as long computation time and restriction of operation types. We challenge to improve computation speed and construct a new application/protocol over homomorphic encryption.
Differential privacy is to randomize data or results, for example adding noise to data or results, to make data unpredictable. Disadvantage of differential privacy is low precision of results. We challenge to compute as high precise results as possible while preserving privacy of original data.
We implemented a national project which is to improve performance (1,000x) of homomorphic encryption from aspects of both computer architecture and theory from 2015 to 2021. We collaborated with other research institutes and archived 2,912 times improvement (26 times by theory and 112 times by middle ware). We also released 13 practical open source libraries.
Privacy-Preserving Deep Learning using Homomorphic Encryption
In the big data era, cloud-based machine learning as a service (MLaaS) has attracted considerable attention. However, when handling sensitive data, such as financial and medical data, a privacy issue emerges, because the cloud server can access clients’ raw data. Therefore, there is a need for both privacy protection of such sensitive data and utilization of the data (e.g. application of machine learning). In this research, we focus on performing the inference process of convolutional neural network (CNN) securely using homomorphic encryption which allow computation over encrypted data without decryption.
Privacy-Preserving Query Response System
In this research, we are working on constructing the privacy-preserving query response system combining homomorphic encryption and differential privacy. Our proposed system protects data owned by Data Providers against both Cloud Servers and Data Analyst by applying differential privacy over ciphertext data using homomorphic operations. In addition, by decrypting differentially private ciphertext data and constructing Differentially Private Database represented in plaintext in advance, our proposed system speeds up the response time to Data Analyst’s queries.
Fully Homomorphic Encryption with Table Lookup for Privacy-Preserving Smart Grid
Smart grid is one of application in smart connected communities (SCC). In order to construct privacy-preserving anomaly detection system on smart grid, we adopt fully homomorphic encryption (FHE) to protect users’ sensitive data. FHE is an encryption scheme that allows a third party to evaluate arbitrary functions using modular arithmetic over encrypted data without decryption. To evaluate any function with FHE, we implemented a protocol replace the calculation to table lookup with FHE. Store the computed function lookup table in cloud server, by searching the input value and select the output value from lookup table with homomorphic encryption, we can decrease the computation cost and run time.
Privacy Preserving Data Falsification Detection in Smart Grids using Elliptic Curve Cryptography and Homomorphic Encryption
Homomorphic Encryption (HE) is a form of encryption that permits users to perform computations on the encrypted data without having to decrypt the data. However, the downside of HE is computational overhead in terms of execution time. Therefore, a method for privacy-preserving and attack detection of data generated by smart meters to shorten the execution time is necessary. Elliptic curve cryptography (ECC) based HE for anomaly-based attack detection for data falsification over encrypted data provides faster computation and ensures data security. Through ECC, we can achieve the same security as a 3,072-bit RSA key with a 256-bit ECC key and HE allows to perform computations on encrypted data. Therefore, ECC based HE requires less memory space to implement the encryption and decryption algorithms, which in turn reduces the time required to perform encryption and decryption operations.
DAMCREM (Dynamic Allocation Method of Computation REsource to Macro-Task)
This is a method named DAMCREM whose goal is decreasing latency of jobs for FHE application over Client-Server model. In DAMCREM, an application is made of macro-tasks which consists of one or more homomorphic operations. DAMCREM aims decreasing latency of jobs by deciding execution order and the number of allocated threads for each macro-task. Detail algorithm of DAMCREM is that application is made of macro-tasks based on types of homomorphic operation. After that, each priority of the number of allocated threads and candidates for each macro-task is made based on measured execution time of each macro-task and each number of threads in advance. During executing the application, DAMCREM schedules execution order of waiting macro-tasks, then decide the number of allocated threads from the candidates for each of the macro-tasks. After that DAMCREM executes the macro-tasks. Previous methods allocates the fix number of threads to each homomorphic operation. Therefore, latency of jobs depends on load of the computation server. DAMCREM decides the number of allocated threads based on the load of the computation server. Thus, DAMCREM can keep low latency of jobs for several load of the computation server.
Large / small comparison operation on ciphertext with reusability
Comparing two homomorphically encrypted integers and its subsequent computation typically require bitwise encryption as it supports arbitrary computation. However, bitwise evaluations for arithmetic circuits such as addition and multiplication over integer are expensive. In this study, we proposed a construction without using bitwise encryption scheme, that is compatible with addition and multiplication of polynomial ring over the encrypted comparison result. We also showed that the depth required for our comparison circuit is the lowest among the previous methods.
Optimization of bootstrapping placement on the compute circuit
Leveled homomorphic encryption, which is a type of fully homomorphic encryption, requires a time-consuming process called bootstrapping each time a certain number of operations between ciphertexts are performed. Here, when the operation you want to calculate in advance is expressed in the form of a circuit, you can reduce the number of times bootstrapping is required for the entire calculation by properly selecting when to execute bootstrapping on the circuit. You need to find the right placement within multiple constraints.
Ring-LWE-based FHE schemes have a problematic feature: ciphertext size increases with every homomorphic multiplication. Furthermore, the computation cost of homomorphic multiplication linearly increases with increasing input sizes. To overcome this, these FHE schemes support a special operation called relinearization, which can reduce the ciphertext size. Relinearization requires almost the same amount of computation cost as that of the homomorphic multiplication, which takes a few to hundreds of milliseconds. Thus, determining when and the number of times to relinearize a ciphertext in a given arithmetic circuit to evaluate in order to minimize the total computation cost is an important task. This problem has been proved to be an NP-hard problem and is called the relinearize problem. In this study, we design an approximation algorithm to address the relinearize problem. The algorithm runs in polynomial time, and we experimentally confirmed that the output of the algorithm is nearly the same as the optimal solution. We also show that the output is exactly equal to the optimal one in a specific case.
Privacy-Preserving Data Classification
Machine learning classification has wide range of applications. In the big data era, a client may need to classify a large amount of data with many features, resulting in heavy computation at the client. Using a cloud server to outsource such classification tasks, we can reduce this computational burden. By applying FHE to classification, the client can outsource classification tasks to a cloud server without revealing any data. In this work, focus on devising a secure classification protocol in which we preserve the privacy of the classification model, the client’s data, and the result while outsourcing computation to a cloud server. In our scenario, the cloud does not learn anything about the classification model, client’s data or result, and the client learns only the result.
Aprili using fully homomorphic encryption (frequent item set mining)
For example, when a pharmaceutical company wants to investigate drugs with frequent side effects from a large number of drug groups, it wants to outsource the calculation, but because the frequency of side effects for each drug is a company secret, it wants to avoid information leakage. The point is how to build a protocol while maintaining the consistency of the decryption result within various restrictions.
Confidential genome search using fully homomorphic encryption
Outsourced private set intersection cardinality using fully homomorphic encryption
Privacy-Preserving Recommendation for Location-Based Service
Extracting useful information by big data analytics such as artificial intelligence technology is receiving a lot of attention from the world. The world is not fully aware of the importance of security, except for banks and hospitals which need high-security assurance levels. Currently, the location-based services such as google map and tabelog hold huge amount of users’ private information. If such huge amount of data are stored and analyzed without encryption, it could lead to critical social problems in the near future. Therefore, this research is about a new mechanism that enables service providers to provide high-quality recommendations without knowing users’ location information and search contents.