Assessing trustworthiness of neural networks for computer systems

Chaudhary, Isha

Permalink

https://hdl.handle.net/2142/129192 Copy

Description

Title

Assessing trustworthiness of neural networks for computer systems

Author(s)

Chaudhary, Isha

Issue Date

2025-04-11

Director of Research (if dissertation) or Advisor (if thesis)

Singh, Gagandeep

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

Neural Networks, Computer Systems, Formal Methods

Abstract

Neural networks have exhibited superior performance in several domains such as computer vision, natural language processing, etc. They have also been trained as computer system components for performance improvement. Although they deliver expected performance, they are not trusted by domain experts, as they are black-box functions and can show unexpected behaviors. This is particularly important for computer systems, as any unpredictable errors in one component can lead to catastrophic failure of the entire system, which is often deployed in consumer-facing applications. Hence, it becomes imperative to develop custom methods to evaluate and develop trust in neural network components of computer systems. In this thesis, I will describe a couple of frameworks that I have developed to address the issue of trust of neural networks in computer systems. The first framework, COMET, is an explanation framework to generate faithful explanations for the predictions of neural networks deployed as performance/cost models in compilers. Cost models statically predict the cost of execution of a given piece of code on a specific CPU. Their predictions are used to direct the compiler optimizations towards the most effective transformations of the code, hence making them crucial for effective compilation. Recent research has developed neural cost models. However, they are not considered trustworthy, as they are black-boxes with no insight into what led to their predictions. Our framework, COMET, is the first step towards mitigating this concern. COMET generates explanations for the prediction of any given cost model for an input code, as a subset of the features of the code that are sufficient to result in the prediction. With explanations, incorrect behaviors of the black-box cost models can be debugged and the trust of the domain experts can be achieved. The second framework, SpecTRA is an automated specification generation system for neural networks as computer system components. We formulate specification generation as an optimization problem and solve it with observations of expected behaviors. We hypothesize that the traditional (aka reference) algorithms that neural networks replace for higher performance can act as effective proxies for expected correct behaviors of the models, when available. SpecTRA clusters similar observations into compact specifications. We present specifications generated by SpecTRA for neural networks in adaptive bit rate and congestion control algorithms. Our specifications show evidence of being correct and matching intuition. Moreover, we use our specifications to show several previously unknown vulnerabilities in SOTA neural models for computer systems.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129192

Copyright and License Information

Copyright 2025 Isha Chaudhary

Owning Collections