Privacy-Preserving Machine Learning via Secure Multiparty Computation

Summary of research carried out

Designing efficient and robust privacy-preserving machine learning (PPML) solutions in the outsourced computation setting (MLaaS) has become the cornerstone of research lately due to the privacy concerns it entails. These privacy concerns hinder collaboration among multiple devices for training ML models, which is otherwise desirable, to obtain efficient models and achieve reasonable accuracy. Thus, there is need for designing PPML algorithms which ensure that no information about the data-set is leaked during training, other than what is permissible by the algorithm. Likewise, no information about the query or the model is revealed during prediction. Towards this, we design secure and efficient multiparty computation (MPC) [13, 6, 2] protocols that can facilitate PPML. MPC allows for n mutually distrusting parties to perform computations on their private inputs, such that no coalition of t parties, controlled by an adversary, can learn anything other than the output of the computation. We design an efficient PPML framework, Tetrad [10], via MPC for 4 servers (n = 4, t = 1) in server-aided computation model providing the strongest security of guaranteed output delivery (GOD). The outsourced computation paradigm is widely being deployed by tech giants such as Amazon, Google, and Microsoft to provide machine learning as a service (MLaaS). The protocols in Tetrad improve upon the state of the art protocols 4-party protocols in [8]. We additionally consider training for deep NN such as LeNet [11] and VGG16 [12] in Tetrad, which was missing in [8]. Tetrad is published at NDSS 2022, a top-tier security conference (rank A). Most GOD protocols in the literature [5, 3, 4, 8, 10] rely on an honest party identified as the trusted third party (TTP) to carry out the computation if misbehavior is detected. Elaborately, the parties entrust the TTP with their inputs, which carries out the computation and delivers the output to all. According to the standard security definition, this leakage of inputs towards a TTP is not considered a privacy breach. However, entrusting a TTP with all the inputs may not be acceptable in real-world applications as this defeats the purpose of employing an MPC protocol. Another drawback of traditional MPC protocols is the view leakage attack. While executing an MPC protocol, nothing prevents an adversary from sending its view to an honest party. This is not treated as an attack in the traditional security definition, since an honest party is expected to discard non-protocol messages, unlike a semi-honest one. However, if this honest party turns rogue in the future, the party can learn sensitive inputs of all parties. This, too, goes against the goal of providing privacy in the system. To address these drawbacks of the traditional MPC security definition, Alon et. al. [1] proposes a new definition, called MPC with Friends and Foes (FaF). This definition requires honest parties’ inputs to be protected against up to t out of n maliciously corrupt parties, and up to h ⋆ semi-honest parties out of the remaining n − t parties. A protocol secure against such an adversary is said to be (t, h⋆ )-FaF secure. Hence, departing from the traditional MPC model, we design FaF-secure MPC protocols. Moreover, keeping efficiency in mind, we design protocols in the 5PC (1, 1)-FaF secure setting. This work, PentaGOD [7], was accepted for publication to ACM CCS 2022, a top-tier conference (rank A). Although the small-party setting has found application in the outsourced computation paradigm, the generic multiparty setting is a better fit for real-world deployments due to its resiliency to a higher number of corruptions (t > 1). Specifically in the honest majority setting, for larger n, the number of corruptions (t < n/2) that can be tolerated is also higher, thereby increasing the trust in the system. Moreover, multiparty setting allows for privacy-conscious computations even in a nonoutsourced deployment scenario, when outsourcing the computation is not feasible/preferable. Hence, we also work towards designing efficient protocols in the honest majority multiparty computation setting (t > 1). This work, MPClan [9], also focuses on designing efficient MPC protocols for PPML, and is currently under submission.

Contributions to CNI

I have been actively helping out with organizing the CNI Networks Seminar Series which is a technical discussion forum in topics in and around computer networks, machine learning, information theory, to name a few. I am a student volunteer and some of my responsibilities have been in helping hosting the talks, identifying and inviting speakers for the seminar series, and following-up with them.

References

[1] Bar Alon, Eran Omri, and Anat Paskin-Cherniavsky. MPC with friends and foes. In CRYPTO, 2020. [2] M. Ben-Or, S. Goldwasser, and A. Wigderson. Completeness Theorems for Non-Cryptographic Fault-Tolerant Distributed Computation (Extended Abstract). In ACM STOC, 1988. [3] Elette Boyle, Niv Gilboa, Yuval Ishai, and Ariel Nof. Practical fully secure three-party computation via sublinear distributed zero-knowledge proofs. In ACM CCS, 2019. [4] Elette Boyle, Niv Gilboa, Yuval Ishai, and Ariel Nof. Efficient fully secure computation via distributed zero-knowledge proofs. In ACM CCS, 2020. [5] M. Byali, A. Joseph, A. Patra, and D. Ravi. Fast secure computation for small population over the internet. In ACM CCS, 2018. [6] O. Goldreich, S. Micali, and A. Wigderson. How to Play any Mental Game or A Completeness Theorem for Protocols with Honest Majority. In STOC, 1987. [7] Nishat Koti, Varsha Bhat Kukkala, Arpita Patra, and Bhavish Raj Gopal. PentaGOD: Stepping beyond traditional GOD with five parties, 2022. [8] Nishat Koti, Mahak Pancholi, Arpita Patra, and Ajith Suresh. SWIFT: Super-fast and Robust Privacy-Preserving Machine Learning. In USENIX, 2021. https://eprint.iacr.org/2020/592. [9] Nishat Koti, Shravani Patil, Arpita Patra, and Ajith Suresh. MPClan: Protocol suite for privacy-conscious computations. [10] Nishat Koti, Arpita Patra, Rahul Rachuri, and Ajith Suresh. Tetrad: Actively secure 4pc for secure training and inference. In NDSS, 2022. https://ia.cr/2021/755. [11] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998. [12] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. [13] A. C. Yao. Protocols for Secure Computations. In FOCS, 1982.