Information Removal at the bottleneck in Deep Neural Networks

Enzo Tartaglione

Abstract Deep neural networks often achieve high performance by exploiting spurious correlations or sensitive attributes present in training data, leading to biased outcomes and privacy concerns. This paper introduces IRENE (Information Removal at the bottleneck), a novel training framework designed to explicitly eliminate specific unwanted information from a model's internal representations. By minimizing a differentiable proxy of mutual information between the network's bottleneck and a sensitive attribute via an adversarial auxiliary head, IRENE effectively "blinds" the model to private features while preserving performance on primary tasks. Experimental results on Biased-MNIST and CelebA demonstrate the method's efficacy and highlight the inherent trade-offs when target and private features are statistically entangled.

1. Introduction

Deep learning models are frequently characterized as "universal problem solvers," yet their tendency to leverage any available statistical signal can be a liability. Models often rely on "spurious correlations"—sensitive attributes such as race, gender, or background artifacts—that happen to be present in the training set. This behavior results in biased decision-making and significant privacy risks if sensitive data can be reconstructed from the model's internal state.

Existing debiasing techniques, such as data re-weighting or balancing, often fail to fully purge this information from the model's latent space. If the information remains present in the "bottleneck" (the final latent representation before classification), it can still be exploited or leaked. In this work, we present IRENE, a method for "pure minimization" of mutual information at the bottleneck, forcing the model to actively unlearn specific features.

2. Methodology

2.1 Structural Architecture

The IRENE framework partitions a standard neural network into three distinct components:

  1. The Encoder ($F$): Processes the input data into a latent representation, or bottleneck ($z$).
  2. The Task Head ($G$): Utilizes the bottleneck $z$ to solve the primary classification task (e.g., identifying if a subject is wearing eyeglasses).
  3. The Information Removal Head ($H$): An auxiliary classifier trained to predict the "private" or sensitive attribute (e.g., gender) from the same bottleneck $z$.

2.2 Gradient Update Strategy

The core innovation of IRENE is its use of a differentiable proxy of Mutual Information (MI). Rather than a standard adversarial loss, which can be unstable, IRENE employs a softmax-based proxy to estimate the information flow between the bottleneck and the private attribute.

The training process functions as a strategic competition:

  • The auxiliary head ($H$) is optimized to maximize its accuracy in predicting the private information from $z$.
  • The encoder ($F$) is simultaneously updated to minimize the mutual information that $H$ is able to extract.
  • Crucially, the gradients from the information removal task are used to update the encoder to make $z$ "un-predictable" for the private attribute, without degrading the auxiliary head's own training stability.

3. Experimental Evaluation

3.1 Datasets

The researchers evaluated IRENE using two benchmarks designed to test information isolation:

  • Biased-MNIST: A synthetic dataset where background colors are correlated with specific digits according to a hyper-parameter $\rho$. This allows for precise control over the degree of bias in the training data.
  • CelebA: A large-scale dataset of celebrity faces. The experiments focused on three primary tasks—identifying Eyeglasses, Blond hair, and Heavy makeup—while attempting to remove Gender as the private information.

3.2 Key Results

In the Biased-MNIST experiments, IRENE successfully maintained the auxiliary head's accuracy at near-random levels (approximately 10% for 10 colors) even when the training correlation was as high as 0.99. This was achieved while maintaining high accuracy on the primary digit classification task.

On the CelebA dataset, the results varied based on the relationship between the target and private features:

  • Disentangled Features: For the "Eyeglasses" task, IRENE reduced gender prediction accuracy from a baseline of ~70% down to ~55% (near the random guess threshold of 50%) while maintaining a 99.68% accuracy on the main task.
  • Entangled Features: For the "Heavy makeup" task, removing gender information caused the main task accuracy to drop from 90.58% to 83.31%. This quantitative shift demonstrates that when features are statistically intertwined, removing one inevitably impacts the model's ability to perceive the other.

4. Discussion and Limitations

IRENE demonstrates that it is possible to "scrub" specific information from a neural network's latent representations. However, the "Entanglement Trap" remains a significant challenge. If a private feature is a fundamental component of how a model perceives a target feature (e.g., the statistical definition of "heavy makeup" in a dataset being heavily tied to "gender"), total erasure of the private feature will degrade performance on the primary task.

5. Conclusion

IRENE provides a robust framework for information removal at the bottleneck, offering a more rigorous approach to privacy and fairness than simple data re-weighting. By integrating information-theoretic concepts directly into the backpropagation flow, it allows for the creation of models that are intentionally blind to sensitive attributes. Future work will explore the application of IRENE to more complex architectures, such as transformers, and its utility in broader debiasing contexts.

Reference: Tartaglione, E. (2022). Information Removal at the bottleneck in Deep Neural Networks. Proceedings of the 33rd British Machine Vision Conference (BMVC).

Summary updated

BibTeX
@inproceedings{tartaglione2022information,
  author = {Enzo Tartaglione},
  title = {{Information Removal at the bottleneck in Deep Neural Networks}},
  booktitle = {BMVC},
  year = {2022},
}
Figure