Fully Homomorphic Encryption (FHE) is a promising solution to enable privacy-preserving inference and training of machine learning models over encrypted data. Among the machine learning methods used in practice, Extreme Gradient Boosting (XGBoost) is one technique that shines in many applications. While previous works have tackled the problem of training tree-based models over FHE, these works either rely on interaction with the client, which adds the extra burden of communication, or consume a typically unreasonable amount of time to train a large model. In this work, we present an efficient system for a non-interactive XGBoost training over FHE that achieves up to 360 imes speedup compared to the state of the art. The argmax operation is a basic building block invoked repeatedly during the XGBoost training as well as other machine learning algorithms, but computing it over FHE is time consuming. When utilizing the Single Instruction Multiple Data (SIMD) parallelism capability offered by most FHE schemes and using a configuration with s slots, the state of the art methods compute argmax on n <= s values using either O(log_2 n) SIMD-comparisons in tournament-style comparison or ceil{n^2 /s} SIMD-comparisons using all pairs comparison. As a second contribution of this work, we propose an efficient argmax algorithm that is based on a novel technique to maximize SIMD-utilization, and computes the argmax of n <= s values using only O(log_2(log_2(n)) SIMD-comparisons. The method extends to n > s with complexity O(n/s) + log_2(log_2(s)), compared to O(n/s) + log_2(s) for state of the art methods. We conduct empirical experiments to compare our method with other existing argmax methods, and show that when using the HEaaN FHE scheme with a configuration of s=2^15 to compute the argmax of n=s values, our implementation is about 1.6 times faster than the state of the art.

Fully homomorphic encryption (FHE) is a powerful cryptographic technique that enables computation on encrypted data without needing to decrypt it. It has broad applications in scenarios where sensitive data needs to be processed in the cloud or in other untrusted environments. FHE applications are both compute- and memory-intensive, owing to expensive operations on large data. While prior works address the challenges of efficient compute using dedicated hardware, expensive memory transfers still remain a major limiting factor. In this work, we propose a hierarchical near-DRAM processing (NDP) solution for FHE applications, called FHENDI, that harnesses the massive DRAM bank bandwidth. We observe various data access patterns in FHE that reveal distinct levels of parallelism: element-wise, limb-wise, coefficient-wise, and ciphertext-wise. FHENDI exploits these levels of parallelism to map FHE operations and data onto different hierarchies of our design, while addressing three major challenges with NDP for FHE: (i) the lack of bank-to-bank communication support, (ii) limited die-to-die bandwidth, and (iii) large memory access latencies. We resolve the first problem through a novel, conflict-free mapping algorithm built atop localized permutation networks that enables efficient element-wise and butterfly operations in FHE. The second problem is addressed by pipelining the execution of parallel bootstrap operations observed in compiled FHE workloads. Finally, we hide the memory access latency behind computation latency by exploiting a dual-banking scheme and subarray-level parallelism (SLP) of the DRAM banks. We evaluate FHENDI using representative workloads in the domains of privacy-preserving machine learning inference on CNNs and Transformers, database range query, and sorting, that are obtained using a compiler framework called HElayers. We compare FHENDI with a server-class CPU and GPU running the state-of-the-art HEaaN library, and an FHE accelerator ASIC, and report mean speedups of $2145.8 \times, 118.29 \times$, and $2.45 \times$, respectively.

Computation delegation to untrusted third-party while maintaining data confidentiality is possible with homomorphic encryption (HE). However, in many cases, the data was encrypted using another cryptographic scheme such as AES-GCM. Hybrid encryption (a.k.a Transciphering) is a technique that allows moving between cryptosystems, which currently has two main drawbacks: 1) lack of standardization or bad performance of symmetric decryption under FHE; 2) lack of input data integrity.

Outsourcing computations over sensitive data to a third-party cloud environment should often rely on dedicated privacy-preserving solutions in order to adhere to privacy regulations such as the GDPR [7]. One solution that gained great attention is fully homomorphic encryption (FHE), a cryptographic method that allows performing different types of computation on encrypted data. Still, writing a non-interactive FHE code that evaluates complex functions is a task that is mostly left to experts. Otherwise, the resulted code may become very slow and even impractical.

Privacy-preserving solutions enable companies to offload confidential data to third-party services while fulfilling their government regulations. To accomplish this, they leverage various cryptographic techniques such as Homomorphic Encryption (HE), which allows performing computation on encrypted data. Most HE schemes work in a SIMD fashion, and the data packing method can dramatically affect the running time and memory costs. Finding a packing method that leads to an optimal performant implementation is a hard task. We present a simple and intuitive framework that abstracts the packing decision for the user. We explain its underlying data structures and optimizer, and propose a novel algorithm for performing 2D convolution operations. We used this framework to implement an inference operation over an encrypted HE-friendly AlexNet neural network with large inputs, which runs in around five minutes, several orders of magnitude faster than other state-of-the-art non-interactive HE solutions.

We present a novel approach for detecting malicious user activity in databases. Specifically, we propose a new machine learning algorithm for detecting attacks such as a stolen user account or illegal use by a user. Our algorithm relies on two main components that examine the consistency of a user's activity and compare it with activity patterns learned from past access. The first component tests for self-consistency, to determine whether the actions performed by a user are consistent with previous patterns. This engine is based on a probabilistic model that we developed to capture a user's normal behavior. The second component checks for global-consistency, to determine whether a user's actions are consistent with the past actions of similar users. We test our algorithm on access data from SQL databases. Experimental results show that we can keep false positive rates while retaining the overall accuracy level. An outlier detection engine based on the presented methods is now included in the standard offering of IBM InfoSphere Guardium1 with positive user feedback.

We use Apache Spark analytics to investigate the logs of an operational cloud object store service to understand how it is being used. This investigation involves going over very large amounts of historical data (PBs of records in some cases) collected over long periods of time retroactively. Existing tools, such as Elasticsearch-Logstash-Kibana (ELK), are mainly used for presenting short-term metrics and can-not perform advanced analytics such as machine learning. A possible solution is to save for long periods only certain aggregations or calculations produced from the raw log data, such as averages or histograms, however these must be decided in advance, and cannot be changed retroactively since the raw data has already been discarded. Spark allows us to gain insights going over historical data collected over long periods of time and to apply the historical models on online data in a simple and efficient way.

Engaging in a debate with oneself or others to take decisions is an integral part of our day-today life.A debate on a topic (say, use of performance enhancing drugs) typically proceeds by one party making an assertion/claim (say, PEDs are bad for health) and then providing an evidence to support the claim (say, a 2006 study shows that PEDs have psychiatric side effects).In this work, we propose the task of automatically detecting such evidences from unstructured text that support a given claim.This task has many practical applications in decision support and persuasion enhancement in a wide range of domains.We first introduce an extensive benchmark data set tailored for this task, which allows training statistical models and assessing their performance.Then, we suggest a system architecture based on supervised learning to address the evidence detection task.Finally, promising experimental results are reported.

The common scenario in computational biology in which a community of researchers conduct multiple statistical tests on one shared database gives rise to the multiple hypothesis testing problem. Conventional procedures for solving this problem control the probability of false discovery by sacrificing some of the power of the tests. We suggest a scheme for controlling false discovery without any power loss by adding new samples for each use of the database and charging the user with the expenses. The crux of the scheme is a carefully crafted pricing system that fairly prices different user requests based on their demands while keeping the probability of false discovery bounded. We demonstrate this idea in the context of HIV treatment research, where multiple researchers conduct tests on a repository of HIV samples.

MOTIVATION: Optimizing HIV therapies is crucial since the virus rapidly develops mutations to evade drug pressure. Recent studies have shown that genotypic information might not be sufficient for the design of therapies and that other clinical and demographical factors may play a role in therapy failure. This study is designed to assess the improvement in prediction achieved when such information is taken into account. We use these factors to generate a prediction engine using a variety of machine learning methods and to determine which clinical conditions are most misleading in terms of predicting the outcome of a therapy. RESULTS: Three different machine learning techniques were used: generative-discriminative method, regression with derived evolutionary features, and regression with a mixture of effects. All three methods had similar performances with an area under the receiver operating characteristic curve (AUC) of 0.77. A set of three similar engines limited to genotypic information only achieved an AUC of 0.75. A straightforward combination of the three engines consistently improves the prediction, with significantly better prediction when the full set of features is employed. The combined engine improves on predictions obtained from an online state-of-the-art resistance interpretation system. Moreover, engines tend to disagree more on the outcome of failure therapies than regarding successful ones. Careful analysis of the differences between the engines revealed those mutations and drugs most closely associated with uncertainty of the therapy outcome. AVAILABILITY: The combined prediction engine will be available from July 2008, see http://engine.euresist.org.

The paper addresses the issue of minimizing the number of nodes involved in routing over a multicast tree and in the maintenance of such a tree in a datagram network. It presents a scheme where the tree routing and maintenance burden is laid only upon the source node and the destination nodes associated with the multicast tree. The main concept behind this scheme is to view each multicast tree as a collection of unicast paths and to locate only the multicast source and destination nodes on the junctions of their multicast tree. The paper shows that despite this restriction, the cost of the created multicast trees is not necessarily higher than the cost of the trees created by other algorithms that do not impose the restriction and therefore require all nodes along the data path of a tree to participate in routing over the tree and in the maintenance of the tree.

The paper addresses the issue of minimizing the number of nodes involved in the routing over a multicast tree and in the maintenance of such a tree in a datagram network. It presents a scheme where the tree routing and maintenance burden is laid only upon the source node and the destination nodes associated with the multicast tree. The main concepts behind this scheme is to view each multicast tree as a collection of unicast paths, and to locate only the multicast source and destination nodes on the junctions of their multicast tree. The paper shows that despite of this restriction, the cost of the created multicast trees is not necessarily higher than the cost of the trees created by other algorithms that do not impose the restriction and therefore require all the nodes along the data path of a tree to participate in the routing over the tree and in the maintenance of the tree.