Towards Reverse-Engineering Black-Box Neural Networks

TL;DR

This paper demonstrates that internal attributes of black-box neural networks can be inferred from query sequences, revealing vulnerabilities and potential for both attacks and defenses, blurring the line between white-box and black-box models.

Contribution

It introduces methods to reverse-engineer internal properties of black-box neural networks from query outputs, highlighting security implications.

Findings

Internal network attributes can be inferred from queries.

Revealed information improves adversarial attack effectiveness.

The technique offers potential for protecting private content.

Abstract

Many deployed learned models are black boxes: given input, returns output. Internal information about the model, such as the architecture, optimisation procedure, or training data, is not disclosed explicitly as it might contain proprietary information or make the system more vulnerable. This work shows that such attributes of neural networks can be exposed from a sequence of queries. This has multiple implications. On the one hand, our work exposes the vulnerability of black-box neural networks to different types of attacks -- we show that the revealed internal information helps generate more effective adversarial examples against the black box model. On the other hand, this technique can be used for better protection of private content from automatic recognition models using adversarial examples. Our paper suggests that it is actually hard to draw a line between white box and black box models.