DEFEND: A Large-scale 1M Dataset and Foundation Model for Tobacco Addiction Prevention

Naga Venkata Sai Raviteja Chappa, Matthew Shepard, Connor McCurtain, Charlotte McCormick, Page Daniel Dobbs, Khoa LuuarXiv 2025 · DOI · arXiv

Introduction

The tobacco industry innovates at an unprecedented pace, frequently introducing new e-cigarettes, nicotine pouches, and sophisticated marketing strategies across digital platforms. Public health regulatory bodies often struggle to keep pace, relying on manual surveillance methods that are increasingly inadequate for this rapidly evolving landscape. The primary technical challenges are two-fold: a severe scarcity of large-scale, high-quality tobacco-related datasets for training robust AI, and the visual complexity of products where brand-specific details and regulatory warnings are critical for classification.

The DEFEND Model: A Dual-Stream Architecture

To address these challenges, researchers have developed DEFEND (Distillation-enabled Enhanced Feature learning for tobacco ENforcement and Discernment). This foundation model utilizes a dual-stream teacher-student architecture designed to capture both global context and fine-grained local details.

The model employs a Teacher (Global Encoder) to process the entire image and a Student (Patch Encoder) to focus on localized patches. Knowledge is distilled from the student to the teacher via Exponential Moving Average (EMA) updates, ensuring comprehensive feature learning. A core innovation is the Feature Enhancement Module (FEM), which uses multi-stage attention to enrich visual and textual features through intra-modal and cross-modal alignment. Furthermore, a specialized local-global visual coherence loss function ensures that fine-grained features remain consistent with the overall product category, while an adaptive sampling strategy focuses the model on distinctive regions like brand identifiers.

The Tobacco-1M Dataset

A significant contribution of this work is the introduction of Tobacco-1M, the first large-scale dataset in this domain. Comprising 1,128,652 high-resolution images, it is approximately 140 times larger than previous datasets. The dataset is organized hierarchically across 75 categories and four levels: Product Type, Usage Context, Content Purpose, and Health Impact.

The Tobacco-1M dataset features a hierarchical labeling system covering combustible, non-combustible, and nicotine replacement products, along with detailed contextual metadata.

The dataset's diversity is extensive, including 450,000 combustible products, 200,000 non-combustibles, and 50,000 nicotine replacement therapies. This scale allows for the training of models that can generalize across the vast variety of tobacco products found globally.

Key Insights and Performance

The research highlights several critical findings:

  • Semantic Regularization: Incorporating detailed textual descriptions—such as mandatory warning labels—significantly improves visual classification. The model learns the underlying semantics of tobacco regulation rather than just visual patterns.
  • Zero-Shot Capabilities: As a foundation model, DEFEND demonstrates a strong ability to identify novel products it has never encountered during training, achieving 45.6% accuracy on unseen categories. This is vital for real-time regulatory monitoring.
  • Precision Attention: Visualization of attention maps confirms that DEFEND focuses on discriminative features like warning labels and brand contours, whereas standard models are often distracted by background noise.

In benchmarking, DEFEND achieved state-of-the-art results, outperforming models like CLIP, CoCa, and DINO. It reached 83.1% Top-1 accuracy on Tobacco-1M and 73.8% accuracy in Visual Question Answering (VQA) tasks regarding product types and health impacts.

Conclusion and Future Directions

DEFEND and the Tobacco-1M dataset provide a powerful new toolkit for public health surveillance. While the model shows impressive zero-shot performance, future work will focus on addressing cultural and language barriers in non-English packaging and expanding the architecture to other domains like alcohol or junk food marketing. This research marks a significant step toward automated, real-time monitoring of harmful product markets.

Summary updated

BibTeX
@misc{chappa2025defend,
  author = {Naga Venkata Sai Raviteja Chappa and Matthew Shepard and Connor McCurtain and Charlotte McCormick and Page Daniel Dobbs and Khoa Luu},
  title = {{DEFEND: A Large-scale 1M Dataset and Foundation Model for Tobacco Addiction Prevention}},
  year = {2025},
  eprint = {2501.13950},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.13950},
}
Figure