Crypto Café at FAU Department of Mathematical Sciences

Our regular Crypto Café seminars take place every other Tuesday 10 am-11 am during the semester. We invite local and international experts on topics in Mathematics and Computer Science related to Cryptography and Information Security.

Come and join us for freshly brewed coffee and interesting conversations on the most exciting topics in cryptography.

Where: SE-43 (Charles E. Schmidt College of Science) - Room 215 and via Zoom

You can catch up on any missed meetings by following the below link:

Upcoming Presentations

December 5, 2023, SE43 - room 215, 10 am    + Zoom (click here)

Speaker:  Dominic Gold, Florida Atlantic University

Title: TDA-Preprocessing Yields Quantifiable Efficiency Gains in Privacy-Preserving ML Models 

Abstract: Computational tools grounded in algebraic topology, known collectively as topological data analysis (TDA), have been used for dimensionality-reduction to preserve salient and discriminating features in data. TDA's flagship method, persistent homology (PH), extracts distinguishing shape characteristics from the data directly and provide inherent noise-tolerance and compact, interpretable representations of high-dimensional data that are amenable to well-established statistical methods and machine learning (ML) models; this faithful but compressed representation of data motivates TDA's use to address the complexity, depth, and inefficiency issues present in privacy-preserving, homomorphic encryption (HE)-based ML models through ciphertext packing---the process of packing multiple encrypted observations into a single ciphertext for Single Instruction, Multiple Data (SIMD) operations.

By investigating several TDA featurization techniques on the MNIST digits dataset using a logistic regression (LR) classifier, we demonstrated that the TDA methods chosen improves encrypted model evaluation with a 10-25 fold reduction in amortized time while improving model accuracy up to 1.4% compared to naive reductions that used downscaling/resizing. The developed technique also has implications for multiclass classification by sending multiple model classifications in a single packed ciphertext to reduce the communication overhead between the Client and Server, potentially avoiding restriction to a binary classification (as done in past HE-ML literature for secure classification of MNIST digits).

Biography: Dominic Gold is a 6th-year graduate teaching assistant at Florida Atlantic University who studies both cryptography and data science, with his main interest in secure/privacy-preserving machine learning on encrypted data. The intersectionality of his research in homomorphic encryption and topological data analysis shows promising implications for research in both fields, with his work in cryptography recognized by venues such as USENIX and ACM CCS. The ultimate goal of his work is to enable real-time predictions on encrypted biomedical data to improve both the reliability, security, and equitability of healthcare systems.

Zoom (click here)

Meeting ID: 878 9825 0483       Passcode: gHJF6g

All are cordially invited.


Recent Presentations

November 21, 2023, SE43 - room 215, 10 am       +Zoom (click here)

Speaker:  Paolo Santini, Universita Polotecnica Delle Marche, Italy

Title: A New Formulation of the Linear Equivalence Problem and Shorter LESS Signatures

Abstract: The problem of determining whether two linear codes are equivalent is called Code Equivalence Problem. When codes are endowed with the Hamming metric (which is the most studied case), the equivalence is mainly considered with respect to monomial transformations (permutations with scaling factors) and the problem is known as the Linear Equivalence Problem (LEP). Code equivalence can be described as a transitive, non-commutative group action and, as such, finds a natural application in cryptography: for example, it is possible to design zero-knowledge proofs, and hence signature schemes. In recent works, it has been shown that LEP can be reformulated using notions such as information sets (arguably, ubiquitous objects in coding theory) and canonical forms. This unlocks some new features, such as the possibility of communicating the equivalence map in a very compact way (which leads to much shorter signatures), as well as opening new attack avenues. In this talk, we recall the basics of code equivalence and then focus on these recent results, aiming to describe how they can be applied to boost the performance of cryptographic schemes.

Zoom (click here)

Meeting ID: 878 9825 0483     Passcode: gHJF6g

Video Recording


November 7, 2023, SE43 - room 215, 10 am       +Zoom (click here)

Speaker: Zhenisbek Assylbekov, Department of Mathematical Sciences, Purdue University Fort Wayne, Fort Wayne, IN

Title: Intractability of Learning AES with Gradient-based Methods

Abstract: We show  the approximate pairwise orthogonality of a class of functions formed by a single AES output bit  under the assumption that all of its round keys except the initial one are independent. This result implies  the hardness of learning AES encryption (and decryption) with gradient-based methods. The proof relies on the Boas-Bellman type of inequality in inner-product spaces.

Keywords: Advanced Encryption Standard, Block Ciphers, Gradient-based Learning

Bio: Zhenisbek has a PhD in Mathematical Statistics from Hiroshima University. After the PhD and some period of work in industry, he got a job at Nazarbayev University, where he was working as a Teaching Assistant, Instructor, and Assistant Professor in the Department of Mathematics during 2011-2023. Currently, he is an Assistant Professor of Data Science at Purdue University Fort Wayne.  His research interests are in machine learning with applications to natural language processing (NLP). He is interested in both the theoretical analysis of machine learning algorithms and the practical implementation and experimental evaluation of such algorithms on text data. He is also interested in hardness of learning which is closely related to cryptography because cryptographic primitives are exactly what is hard for machine learning.

Video Recording