Analyzing and Applying Cryptographic Mechanisms to Protect Privacy in Applications

Privacy-Enhancing Technologies (PETs) emerged as a technology-based response to the increased collection and storage of data as well as the associated threats to individuals' privacy in modern applications. They rely on a variety of cryptographic mechanisms that allow to perform some computation without directly obtaining knowledge of plaintext information. However, many challenges have so far prevented effective real-world usage in many existing applications. For one, some mechanisms leak some information or have been proposed outside of security models established within the cryptographic community, leaving open how effective they are at protecting privacy in various applications. Additionally, a major challenge causing PETs to remain largely academic is their practicality-in both efficiency and usability. Cryptographic mechanisms introduce a lot of overhead, which is mostly prohibitive, and due to a lack of high-level tools are very hard to integrate for outsiders. In this thesis, we move towards making PETs more effective and practical in protecting privacy in numerous applications. We take a two-sided approach of first analyzing the effective security (cryptanalysis) of candidate mechanisms and then building constructions and tools (cryptographic engineering) for practical use in specified emerging applications in the domain of machine learning crucial to modern use cases. In the process, we incorporate an interdisciplinary perspective for analyzing mechanisms and by collaboratively building privacy-preserving architectures with requirements from the application domains' experts. Cryptanalysis. While mechanisms like Homomorphic Encryption (HE) or Secure Multi-Party Computation (SMPC) provably leak no additional information, Encrypted Search Algorithms (ESAs) and Randomization-only Two-Party Computation (RoTPC) possess additional properties that require cryptanalysis to determine effective privacy protection. ESAs allow for search on encrypted data, an important functionality in many applications. Most efficient ESAs possess some form of well-defined information leakage, which is cryptanalyzed via a breadth of so-called leakage attacks proposed in the literature. However, it is difficult to assess their practical effectiveness given that previous evaluations were closed-source, used restricted data, and made assumptions about (among others) the query distribution because real-world query data is very hard to find. For these reasons, we re-implement known leakage attacks in an open-source framework and perform a systematic empirical re-evaluation of them using a variety of new data sources that, for the first time, contain real-world query data. We obtain many more complete and novel results where attacks work much better or much worse than what was expected based on previous evaluations. RoTPC mechanisms require cryptanalysis as they do not rely on established techniques and security models, instead obfuscating messages using only randomizations. A prominent protocol is a privacy-preserving scalar product protocol by Lu et al. (IEEE TPDS'13). We show that this protocol is formally insecure and that this translates to practical insecurity by presenting attacks that even allow to test for certain inputs, making the case for more scrutiny of RoTPC protocols used as PETs. This part of the thesis is based on the following two publications: [KKM+22] S. KAMARA, A. KATI, T. MOATAZ, T. SCHNEIDER, A. TREIBER, M. YONLI. “SoK: Cryptanalysis of Encrypted Search with LEAKER - A framework for LEakage AttacK Evaluation on Real-world data”. In: 7th IEEE European Symposium on Security and Privacy (EuroS&P’22). Full version: https://ia.cr/2021/1035. Code: https://encrypto.de/code/LEAKER. IEEE, 2022, pp. 90–108. Appendix A. [ST20] T. SCHNEIDER , A. TREIBER. “A Comment on Privacy-Preserving Scalar Product Protocols as proposed in “SPOC””. In: IEEE Transactions on Parallel and Distributed Systems (TPDS) 31.3 (2020). Full version: https://arxiv.org/abs/1906.04862. Code: https://encrypto.de/code/SPOCattack, pp. 543–546. CORE Rank A*. Appendix B. Cryptographic Engineering. Given the above results about cryptanalysis, we investigate using the leakage-free and provably-secure cryptographic mechanisms of HE and SMPC to protect privacy in machine learning applications. As much of the cryptographic community has focused on PETs for neural network applications, we focus on two other important applications and models: Speaker recognition and sum product networks. We particularly show the efficiency of our solutions in possible real-world scenarios and provide tools usable for non-domain experts. In speaker recognition, a user's voice data is matched with reference data stored at the service provider. Using HE and SMPC, we build the first privacy-preserving speaker recognition system that includes the state-of-the-art technique of cohort score normalization using cohort pruning via SMPC. Then, we build a privacy-preserving speaker recognition system relying solely on SMPC, which we show outperforms previous solutions based on HE by a factor of up to 4000x. We show that both our solutions comply with specific standards for biometric information protection and, thus, are effective and practical PETs for speaker recognition. Sum Product Networks (SPNs) are noteworthy probabilistic graphical models that-like neural networks-also need efficient methods for privacy-preserving inference as a PET. We present CryptoSPN, which uses SMPC for privacy-preserving inference of SPNs that (due to a combination of machine learning and cryptographic techniques and contrary to most works on neural networks) even hides the network structure. Our implementation is integrated into the prominent SPN framework SPFlow and evaluates medium-sized SPNs within seconds. This part of the thesis is based on the following three publications: [NPT+19] A. NAUTSCH, J. PATINO, A. TREIBER, T. STAFYLAKIS, P. MIZERA, M. TODISCO, T. SCHNEIDER, N. EVANS. Privacy-Preserving Speaker Recognition with Cohort Score Normalisation”. In: 20th Conference of the International Speech Communication Association (INTERSPEECH’19). Online: https://arxiv.org/abs/1907.03454. International Speech Communication Association (ISCA), 2019, pp. 2868–2872. CORE Rank A. Appendix C. [TNK+19] A. TREIBER, A. NAUTSCH , J. KOLBERG , T. SCHNEIDER , C. BUSCH. “Privacy-Preserving PLDA Speaker Verification using Outsourced Secure Computation”. In: Speech Communication 114 (2019). Online: https://encrypto.de/papers/TNKSB19.pdf. Code: https://encrypto.de/code/PrivateASV, pp. 60–71. CORE Rank B. Appendix D. [TMW+20] A. TREIBER , A. MOLINA , C. WEINERT , T. SCHNEIDER , K. KERSTING. “CryptoSPN: Privacy-preserving Sum-Product Network Inference”. In: 24th European Conference on Artificial Intelligence (ECAI’20). Full version: https://arxiv.org/abs/2002.00801. Code: https://encrypto.de/code/CryptoSPN. IOS Press, 2020, pp. 1946–1953. CORE Rank A. Appendix E. Overall, this thesis contributes to a broader security analysis of cryptographic mechanisms and new systems and tools to effectively protect privacy in various sought-after applications