Network Inversion of Convolutional Neural Nets (Student Abstract)

Pirzada Suhail, Amit SethiAAAI 2025 · DOI

Abstract

Deep neural networks are highly effective but notoriously opaque. This paper introduces a novel network inversion technique designed to reconstruct the inputs that trigger specific classifier predictions, thereby improving model interpretability. By utilizing a conditioned generator with vector-matrix label encoding and a multi-part loss function that penalizes feature similarity, the authors successfully visualize diverse internal features across standard benchmarks like MNIST and CIFAR10.

Introduction

The "black box" nature of deep learning poses significant risks in safety-critical applications such as medicine or autonomous driving. Network inversion offers a window into these models by synthesizing inputs that elicit specific responses. However, existing inversion techniques frequently encounter "mode collapse," where the generator produces repetitive, non-diverse samples. This research proposes a more efficient and diverse inversion process to better capture the full breadth of features learned by a network.

Methods

The core of the approach is a conditioned generator that learns the data distribution of a pre-trained classifier's input space. Three technical innovations are key to this framework:

  1. Vector-Matrix Conditioning: Instead of simple class labels, labels are encoded into $n$-dimensional vectors and $n \times n$ "hot matrices." This provides the generator with sufficient variance to explore the input space.
  2. Diversity-Focused Architecture: High dropout rates are applied during the up-convolution (generation) process, forcing the network to discover multiple paths for representing a single class and preventing reliance on "shortcut" solutions.
  3. Multi-Component Inversion Loss ($L_{Inv}$): The training objective balances three components: KL Divergence ($L_{KL}$) for distribution matching, Cross-Entropy ($L_{CE}$) for classification accuracy, and Cosine Similarity ($L_{Cosine}$) to minimize feature overlap between generated samples.
System architecture showing the interaction between the generator, classifier, and the multi-component loss function.

Results

The method was evaluated on four standard computer vision benchmarks: MNIST, FashionMNIST, SVHN, and CIFAR10. Qualitative results indicate that the generator successfully reconstructs recognizable and varied inputs for all target classes. A key finding was that the inclusion of the Cosine Similarity loss significantly enhanced both intra-class diversity (variety within a single label) and inter-class distinction. By leveraging the classifier's inherent robustness, the inversion process can be guided to reconstruct data that resembles the original training set rather than random noise.

Visualizations of inverted features for various datasets, demonstrating the diversity achieved by the proposed method.

Conclusion and Future Work

This work demonstrates that sophisticated conditioning and diversity-promoting loss functions can effectively invert convolutional neural networks, providing a promising step toward opening the "black box" of deep learning. While current results are primarily qualitative, future efforts will involve developing quantitative metrics and testing the scalability of this vector-matrix conditioning on higher-resolution datasets like ImageNet.

Summary updated

BibTeX
@inproceedings{suhail2025network,
  author = {Pirzada Suhail and Amit Sethi},
  title = {{Network Inversion of Convolutional Neural Nets (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35303},
}
Figure