Due to the emerging privacy-protection laws and regulations (e.g. GDPR in the EU) in recent years, dozens of multi-party computation (MPC for short) protocols have been proposed and widely applied by companies and institutions. These MPC protocols enable companies and institutions to perform joint analyses and machine learning on their private data while protecting their data's privacy. However, due to the complexity of MPC protocols, their implementations of-ten contain data leakage vulnerabilities, which can critically undermine the intended privacy protection. Additionally, most existing security analyses of MPC protocols rely on theoretical proofs, neglecting to detect possible vulnerabilities in MPC im-plementations. Therefore, detecting data leakage vulnerabilities in MPC implementations is an urgent necessity. In this paper, we propose MPCGuard, a practical frame-work for detecting data leakage vulnerabilities in MPC imple-mentations. Different from traditional memory vulnerabilities, data leakage vulnerabilities in MPC implementations cannot be identified by existing sanitizers. To resolve this challenge, we first establish a leakage identifier in MPCGuard with two neural network classifiers to identify whether an MPC implementation contains data leakage vulnerabilities. To enhance identification effectiveness, the structures of neural network classifiers are designed according to the characteristics of MPC protocols. After identifying a data leakage vulnerability, we employ a delta method to assist in locating the vulnerability. To demonstrate the effectiveness of MPCGuard, we apply MPCGuard to test 29 commonly-used MPC implementations in three main-stream MPC frameworks, i.e. Crypten, TF-Encrypted, and MP-SPDZ. We discover that 12 out of 29 implementations contain data leakage vulnerabilities, some of which can lead to the reconstruction of raw data. Until the moment this paper is written, all vulnerabilities, two of which have been assigned with CVE-IDs, have been confirmed. To the best of our knowledge, these two CVE-IDs are the first CVE-IDs assigned for data leakage vulnerabilities in MPC implementations.

Due to the rising privacy concerns on sensitive client data and trained models like Transformers, secure multi-party computation (MPC) techniques are employed to enable secure inference despite attendant overhead. Existing works attempt to reduce the overhead using more MPC-friendly non-linear function approximations. However, the integration of quantization widely used in plaintext inference into the MPC domain remains unclear. To bridge this gap, we propose the framework named Ditto to enable more efficient quantization-aware secure Transformer inference. Concretely, we first incorporate an MPC-friendly quantization into Transformer inference and employ a quantization-aware distillation procedure to maintain the model utility. Then, we propose novel MPC primitives to support the type conversions that are essential in quantization and implement the quantization-aware MPC execution of secure quantized inference. This approach significantly decreases both computation and communication overhead, leading to improvements in overall efficiency. We conduct extensive experiments on Bert and GPT2 models to evaluate the performance of Ditto. The results demonstrate that Ditto is about $3.14\sim 4.40\times$ faster than MPCFormer (ICLR 2023) and $1.44\sim 2.35\times$ faster than the state-of-the-art work PUMA with negligible utility degradation.

In the business scenarios at Ant Group, there is a rising demand for collaborative data analysis among multiple institutions, which can promote health insurance, financial services, risk control, and others. However, the increasing concern about privacy issues has led to data silos. Secure Multi-Party Computation (MPC) provides an effective solution for collaborative data analysis, which can utilize data value while ensuring data security. Nevertheless, the performance bottlenecks of MPC and the strong demand for scalability pose great challenges to secure collaborative data analysis frameworks. In this paper, we build a secure collaborative data analysis system SCQL with a general purpose. We design more efficient MPC protocols and relational operators to meet the demand for scalability. In terms of system design, we aim to implement a system with security, usability, and efficiency. We conduct extensive experiments on SCQL to validate our optimization improvements: (1) Our optimized secure sort protocol sorts one million 64-bit data in only 4.5 minutes, 126× faster than EMP (9.4 hours). (2) The end-to-end execution time of the typical vertical scenario query is reduced by 1991× from the state-of-the-art semi-honest collaborative analysis framework Secrecy (rewritten with Additive Secret Sharing protocol), with appropriate security tradeoffs. (3) We test the system in the WAN setting with input size = 10 7 to demonstrate the scalability. We have successfully deployed SCQL to address problems in real-world business scenarios at Ant Group.

As the security issue of cloud-offloaded Deep Learning (DL) inference is drawing increasing attention, running DL inference in Trusted Execution Environments (TEEs) has become a common practice. Latency prediction of TEE-hosted DL model inference is essential for many scenarios, such as DNN model architecture searching with a latency constraint or layer scheduling in model-parallelism inference. However, existing solutions fail to address the memory over-commitment issue in resource-constrained environments inside TEEs.This paper presents Sectum, an accurate latency predictor for DL inference inside TEE enclaves. We first perform a synthetic empirical study to analyze the relationship between inference latency and memory occupation. Sectum predicts inference latency following a two-stage design based on some critical observations. First, Sectum uses a Graph Neural Network (GNN)-based model to detect whether a given model would trigger memory over-commitment in TEEs. Then, combining operator-level latency modeling with linear regression, Sectum could predict the latency of a model. To evaluate Sectum, we design a large dataset that contains the latency information of over 6k CNN models. Our experiments demonstrate that Sectum could achieve over 85% ±10% accuracy of latency prediction. To our knowledge, Sectum is the first method to predict TEE-hosted DL inference latency accurately.

Hyper-parameter tuning (HPT) is crucial for many machine learning (ML) algorithms. But due to the large searching space, HPT is usually time-consuming and resource-intensive. Nowadays, many researchers use public cloud resources to train machine learning models, convenient yet expensive. How to speed up the HPT process while at the same time reduce cost is very important for cloud ML users. In this paper, we propose SpotTune, an approach that exploits transient revocable resources in the public cloud with some tailored strategies to do HPT in a parallel and cost-efficient manner. Orchestrating the HPT process upon transient servers, SpotTune uses two main techniques, fine-grained cost-aware resource provisioning, and ML training trend predicting, to reduce the monetary cost and runtime of HPT processes. Our evaluations show that SpotTune can reduce the cost by up to 90% and achieve a 16.61x performance-cost rate improvement.

EMR (Elastic Map Reduce) is a service provided by mainstream cloud vendors for data processing users to directly obtain well-managed Hadoop YARN clusters on the cloud. Preemptible instance is a kind of cloud server that is cheap but is likely to be reclaimed by cloud vendors suddenly. Running EMR clusters on preemptible instances relies on YARN's own fault-tolerance, which is limited. In this paper, we present PrTaurus as an availability-enhanced EMR service on preemptible instances. PrTaurus integrates a system-level checkpoint capability based on Docker into YARN to further improve its fault-tolerance. In addition, PrTaurus's scheduling strategy takes advantage of Alibaba Cloud's one-hour protection policy. Furthermore, a new method that comprehensively considers cost-efficiency, preemption risk and overhead is proposed to select cluster instances. We evaluated PrTaurus through simulations on real-world workload and instance price traces. Experimental results show that compared with the existing EMR clusters running on preemptible instances, PrTaurus significantly reduces cost (13.0%-74.6%), instance preemptions (60.3%-88.9%), and task preemptions (86.0 % - 98.6 %).

Modern cloud vendors provide thousands of types of virtual machines, presenting the user a bewildering choice. Many previous works try to train accurate models for a workload to predict its performance on a given configuration, which could be applicable to a concrete class of workloads (e.g., Spark-SVM jobs). But it is difficult to build a universal prediction model for all the workloads, mostly because different workloads have different runtime metrics distribution. We argue that if the runtime metrics data could be transferred before being fed to the prediction model, the knowledge learned in one workload may be reused to other workloads. In this paper, we synthetically investigate the similarity and difference between a set of workloads and the possibility to transfer between them. We use an LSTM auto-encoder to encode the temporal runtime metrics into a fixed-size vector, which is fed into a linear regression model. In our experiments, the metrics transferring approach could help reduce the prediction error of the linear regression model.

Abstract Currently the generative adversarial networks (GANs) have rapidly become a popular research hotspot that people concerned and have been applied to various fields. Lots of meaningful work have been proposed and various variants of GANs sprung up in last few years. The scholars usually design GAN structure like the layers and hyperparameters setting according to the experience and constantly attempts. For the propose of finding the appropriate structure more conveniently and efficiently. A method with multiobjective algorithm is proposed to obtain the optimal structure for the GANs. In the proposed method, the nondominated sorting genetic algorithm II (NSGA‐II) is utilized to optimize the hyperparameters and structure of deep convolutional generative adversarial network (DCGAN). The experiments are conducted on MNIST and Malware datasets demonstrate the efficiency and high performance of proposed method.

The increasing popularity of cloud storage is leading many organizations to move their data into the cloud. However, putting all data in one cloud causes problems such as vendor lock-in, increased service costs, and data availability. In this paper, we introduce DCStore, a Cloud-of-Clouds storage service designed for an organization to outsource their data into the clouds. To achieve the goal of cost-efficient and high-available, we combine three key techniques. First, DCStore eliminates the redundant data at client-side to save storage cost via application-aware chunking method. Second, DCStore uses an inner-chunk based erasure coding scheme to distribute unique chunks across multiple clouds for high availability. Finally, a container-based share management strategy is used for performance optimization. Our experimental evaluations show that DCStore can improve the performance and cost efficiency significantly, compared with existing Cloud-of-Clouds storage systems.

The size of data sets being collected and analyzed in data science field is growing rapidly, making traditional big data processing solution prohibitively expensive. Especially when the data sets are too large, distributed techniques are inevitable even for simple embarrassing parallel jobs. However, distributed computing is still inaccessible to a large number of users. For example, many average users are still struggling with complex cluster management and configuration tools[24] even just for summing up a group of numbers in a large data file. In this paper, we present BDViewer, a web-based big data processing and visualizing tool. BDViewer uses JavaScript plugins to enable users to view, process and visualize their large data files just through a web browser. By just clicking a button, users can open a large data file online and view the file contents immediately no matter how large the file is. In the back-end, BDViewer is built on a virtual private cloud system. Users' operations in a web browser are converted into map-reduce jobs and MPI tasks that are executed on the cloud. At the end of this paper, some experiments are carried out, which demonstrate BDViewer's effectiveness and ease of use.

The emergence of cloud computing makes it easier for enterprises and individual users to obtain distributed computing resources. At the same time, developing distributed systems is still complex and difficult. In order to reduce the complexity in distributed application development, a series of techniques such as RPC, message-oriented middleware have been proposed and widely used. These techniques hide the underlying communicating details but require developers to take control of coordination among different machines for complex distributed tasks. This paper proposes a lightweight event-driven framework called DistGear to simplify the development of complex task processing in distributed applications. DistGear is a layer of software built on top of communication facilities for providing developers with a programming abstraction to easily handle distributed tasks and control machine coordination in tasks. We implement DistGear based on coroutines and message communication. Experience with DistGear in a simplified real-world system suggests it is effective. Performance evaluation also demonstrates DistGear's good concurrency and low resources consumption.