Abstract
Generative adversarial networks (GANs) are a framework that learns a generative distribution through adversarial training. Recently, their class conditional extensions (e.g., conditional GAN (cGAN) and auxiliary classifier GAN (AC-GAN)) have attracted much attention owing to their ability to learn the disentangled representations and to improve the training stability. However, their training requires the availability of large-scale accurate class-labeled data, which are often laborious or impractical to collect in a real-world scenario. To remedy this, we propose a novel family of GANs called label-noise robust GANs (rGANs), which, by incorporating a noise transition model, can learn a clean label conditional generative distribution even when training labels are noisy. In particular, we propose two variants: rAC-GAN, which is a bridging model between AC-GAN and the label-noise robust classification model, and rcGAN, which is an extension of cGAN and solves this problem with no reliance on any classifier. In addition to providing the theoretical background, we demonstrate the effectiveness of our models through extensive experiments using diverse GAN configurations, various noise settings, and multiple evaluation metrics (in which we tested 402 conditions in total).
1. Introduction
Conditional Generative Adversarial Networks (GANs), such as cGAN and AC-GAN, have significantly advanced the field of image synthesis by allowing for class-specific generation. However, these models are highly sensitive to the quality of training labels. In practical applications, datasets often suffer from "label noise"—incorrect or misleading labels arising from automated data collection or human error. Standard conditional GANs typically fail in these environments, leading to poor image quality and a failure to disentangle distinct classes, a phenomenon often described as "garbage in, garbage out."
This paper introduces Label-Noise Robust GANs (rGANs), a novel framework that enables conditional generative models to learn clean image distributions from noisy datasets. By integrating a noise transition model directly into the GAN architecture, the authors demonstrate that it is possible to generate high-quality, accurately categorized images even when a vast majority of the training labels are corrupted.
2. Methodology
The rGAN framework introduces two robust variants based on the most prevalent conditional GAN architectures. The central component of both methods is a Noise Transition Model, represented by a matrix $T$, where $T_{i,j}$ denotes the probability that a true label $i$ is observed as a noisy label $j$. The authors provide theoretical proof that if $T$ is non-singular, the generator can converge to the true, clean data distribution.
2.1 rAC-GAN (Robust Auxiliary Classifier GAN)
Standard AC-GANs use an auxiliary classifier to guide the generator. rAC-GAN enhances this by applying a "forward correction" technique. It assumes the existence of a latent clean label that is flipped to a noisy label according to the transition matrix $T$. The model is trained to minimize a loss function that accounts for this flipping probability, effectively filtering the noise during the learning process.
2.2 rcGAN (Robust Conditional GAN)
In contrast to AC-GAN, rcGAN extends the standard cGAN architecture, which feeds labels directly into the discriminator. In rcGAN, the generator produces an image based on a clean label, but this label is intentionally "corrupted" using the noise transition model before being presented to the discriminator. This alignment ensures the generated distribution matches the noisy distribution of the real data, forcing the generator to recover the underlying clean mapping.
3. Experimental Evaluation
The researchers evaluated rGANs across 402 experimental conditions, utilizing benchmarks such as CIFAR-10, CIFAR-100, and the real-world Clothing1M dataset, which contains naturally occurring noise from online metadata.
3.1 Results and Key Insights
As illustrated in Figure 1, standard cGANs struggle to maintain class separation when trained on noisy data, often mixing features from different categories. In contrast, the proposed rcGAN successfully recovers the clean distribution, producing distinct and high-quality samples for each class.
Quantitative results using Fréchet Inception Distance (FID) and GAN-train/test metrics confirm that rGANs consistently outperform baselines. A notable observation is the "memorization effect": while cGANs tend to learn the noisy distribution directly, AC-GANs exhibit slight inherent robustness because their internal classifiers prioritize simple, clean patterns before memorizing noise. For extreme noise scenarios (e.g., 90% corruption), the authors found that adding a mutual information regularization term ($L_{MI}$) is essential to prevent the generator from ignoring the labels entirely.
4. Conclusion
Label-Noise Robust GANs provide a robust mathematical framework for training conditional generative models on imperfectly labeled data. While the current model assumes class-dependent noise and performs best when the transition matrix $T$ is known or accurately estimated, it represents a significant advancement toward practical, large-scale generative modeling. Future research may focus on feature-dependent noise models and automated methods for estimating transition matrices in highly adversarial environments.
Summary updated