Adversary-Aware Partial label learning with Label distillation

Cheng Chen, Yueming Lyu, Ivor W. Tsang

TL;DR

This paper introduces a novel privacy-preserving framework called Adversary-Aware Partial Label Learning (PLL), which intentionally injects "rival labels" into training data to protect sensitive user information from data breaches. To overcome the massive label noise and mathematical intractability this encryption creates, the authors propose the Adversarial Teacher within Momentum (ATM) algorithm. ATM successfully disambiguates the true labels, delivering a provably consistent classifier that outperforms existing state-of-the-art methods on standard image benchmarks.

What Problem Does This Paper Address?

Partial Label Learning (PLL) is a branch of weakly supervised learning where each training instance is assigned a set of candidate labels, only one of which is the ground truth. While PLL reduces annotation costs, it suffers from a major vulnerability: data privacy. If a partially labeled dataset collected from human subjects is leaked or stolen, an adversary can use standard PLL algorithms to decode the true labels and extract confidential information.

To mitigate this, one can intentionally corrupt the dataset by adding "rival labels" (fake candidate labels) to act as a form of encryption. However, this defense mechanism introduces a severe side effect. It creates an exceptionally high ratio of noisy labels, making the mathematical "transition matrix" (which maps true labels to noisy candidate sets) completely intractable. Standard PLL methods break down under this level of ambiguity, resulting in severe classifier inconsistency and poor performance.

Methods

The authors tackle this challenge by co-designing a novel label generation process and a robust optimization algorithm:

  1. Adversary-Aware Label Generation: They extend standard PLL by factorizing the label generation probability to include a controllable rival label ($Y'$). This acts as local differential privacy, providing a tunable knob ($\bar{T}$) to balance privacy protection against learning difficulty.
  2. Adversarial Teacher within Momentum (ATM): To handle the intractable transition matrix ($Q^*$), the authors introduce a MoCo-style momentum dictionary technique. Instead of explicitly calculating the massive matrix, ATM leverages self-attention prototype learning to generate soft pseudo-labels from instance embeddings, drastically reducing matrix uncertainty.
  3. Dual-Term Loss Function: The model is trained using a specialized loss function (Equation 19) consisting of:
  4. A debiased cross-entropy loss that uses the known independent transition matrix to down-weight false positive predictions.
  5. A contrastive loss that maximizes the similarity between an instance and a dynamically selected positive sample set, boosting representation learning.
Figure 1: An overview of the proposed method. General partial label can be disclosed to adversary. The initial training is about positive sample selection. Moreover, we have assumed ¯T is given.

Datasets

The proposed method was evaluated by manually corrupting three widely used image classification benchmarks to simulate an adversary-aware environment:

  • CIFAR10
  • CIFAR100
  • CUB200 (a fine-grained dataset of bird species)

Key Insights and Ideas

  • Privacy via Intentional Corruption: Rather than treating label noise as a nuisance to be avoided, this work treats structured noise ("rivals") as a deliberate feature for data privacy.
  • Decoupling the Transition Matrix: By assuming that rival label generation depends only on the true label and not on the complex instance features, the authors make the class-independent transition matrix mathematically identifiable and cheap to implement as an encryption mechanism.
  • Virtuous Cycle of Disambiguation: The debiased loss function helps identify cleaner positive samples for contrastive learning. In turn, better contrastive representations enhance the accuracy of the self-attention prototypes, creating a feedback loop that steadily eliminates label ambiguity.

Key Results

The ATM framework demonstrated remarkable resilience in highly ambiguous and adversarial settings, consistently outperforming top-tier PLL methods like PiCO, LWS, and PRODEN:

  • High Noise Resilience: On the CIFAR10 dataset with a high partial label rate ($q = 0.5$), the proposed method achieved an 8.17% performance increase over the previous state-of-the-art method (PiCO).
  • Adversarial Superiority: Under the strict adversary-aware setting ($q^* = 0.1$), ATM achieved 54.09% accuracy on CIFAR100 and 56.26% accuracy on CUB200, outperforming PiCO by roughly 6% on both datasets, while baseline methods like LWS completely collapsed (dropping to <5% accuracy).

Limitations and Future Work

While the class-independent transition matrix assumption makes the problem tractable and affordable to deploy, real-world adversarial attacks might exploit instance-dependent vulnerabilities. Additionally, approximating the transition matrix in massive label spaces remains computationally intensive.

For future work, the authors suggest exploring variational inference methods to better approximate the intractable transition matrix without relying on rigid structural assumptions.

Summary updated

Abstract

To ensure that the data collected from human subjects is entrusted with a secret, rival labels are introduced to conceal the information provided by the participants on purpose. The corresponding learning task can be formulated as a noisy partial-label learning problem. However, conventional partial-label learning (PLL) methods are still vulnerable to the high ratio of noisy partial labels, especially in a large labelling space. To learn a more robust model, we present Adversary-Aware Partial Label Learning and introduce the $\textit{rival}$, a set of noisy labels, to the collection of candidate labels for each instance. By introducing the rival label, the predictive distribution of PLL is factorised such that a handy predictive label is achieved with less uncertainty coming from the transition matrix, assuming the rival generation process is known. Nonetheless, the predictive accuracy is still insufficient to produce an sufficiently accurate positive sample set to leverage the clustering effect of the contrastive loss function. Moreover, the inclusion of rivals also brings an inconsistency issue for the classifier and risk function due to the intractability of the transition matrix. Consequently, an adversarial teacher within momentum (ATM) disambiguation algorithm is proposed to cope with the situation, allowing us to obtain a provably consistent classifier and risk function. In addition, our method has shown high resiliency to the choice of the label noise transition matrix. Extensive experiments demonstrate that our method achieves promising results on the CIFAR10, CIFAR100 and CUB200 datasets.
BibTeX
@misc{chen2023adversary,
  author = {Cheng Chen and Yueming Lyu and Ivor W. Tsang},
  title = {{Adversary-Aware Partial label learning with Label distillation}},
  year = {2023},
  eprint = {2304.00498},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.00498},
}
Figure