A privacy-preserving approach for cloud-based protein fold recognition

Abstract

The complexity and cost of training machine learning models have made cloud-based machine learning as a service (MLaaS) attractive for businesses and researchers. MLaaS eliminates the need for in-house expertise by providing pre-built models and infrastructure. However, it raises data privacy and model security concerns, especially in medical fields like protein fold recognition. We propose a secure three-party computation-based MLaaS solution for privacy-preserving protein fold recognition, protecting both sequence and model privacy. Our efficient private building blocks enable complex operations privately, including addition, multiplication, multiplexer with a different methodology, most-significant bit, modulus conversion, and exact exponential operations. We demonstrate our privacy-preserving recurrent kernel network (RKN) solution, showing that it matches the performance of non-private models. Our scalability analysis indicates linear scalability with RKN parameters, making it viable for real-world deployment. This solution holds promise for converting other medical domain machine learning algorithms to privacy-preserving MLaaS using our building blocks.

Keywords: cloud-based machine learning; data privacy; machine learning as a service; multi-party computation; privacy preserving machine learning; protein fold recognition; recurrent kernel networks.

Figure 1

The overview of our privacy-preserving RKN as a service via MPC (1) At first, the data owner, i.e., Alice, secret shares her data and sends the shares to the proxies. Similarly, the model owner, i.e., Bob, does the same with the parameters of the model. (2) Then, by using the outsourced model and the data, the proxies, $P_0$ and $P_1$, perform the operations required for the inference of the data on the model with the help of $P_2$, which is the helper. (3) Finally, the proxies send the shares of the prediction of the given data to the data owner.

Figure 2

The architecture of RKN and internal computations of its layers (A and B) The arithmetic circuits of (A) a single neuron of RKN at position t and k-mer level u and (B) the linear classifier layer of RKN after the last position of the input sequence x are depicted. $z_i^j$ represents the i-th character of the j-th anchor point, and $c_i\left[j\right]$ represents the initial mapping of the sequence up to its j-th character into a q-dimensional vector based on anchor points of length i where $i\in \left\{1,\dots ,k\right\}$, $j\in \left\{1,\dots ,s\right\}$, and q is the number of anchor points. (C) For $k=3$ and $\left|x\right|=5$, RKN is shown where the green nodes are the single neurons and the pink one is the linear classifier.

Figure 3

The results of the execution time analysis of our RKN as a service The results of the execution time analysis of our RKN as a service on both WAN and LAN settings for varying (A) numbers of anchor points for a fixed k-mer length and sequence length, (B) lengths of k-mers for a fixed number of anchor points and sequence length, and (C) lengths of sequences for a fixed number of anchor points and k-mer length.