How Alibaba’s AI Powers Brand Risk Detection: Models, Data, and Results

Background

Brand risk detection is essential for advertising platforms to prevent the sale of counterfeit or unlicensed goods. The system focuses on visual brand identification, which must handle diverse product categories, varying angles, occlusions, tiny or missing logos, and strong adversarial attacks.

Challenges

High visual diversity across categories and styles.

Multiple brand‑specific features (logos, patterns, designs).

Logos may be very small or absent.

Large number of brands with a highly imbalanced data distribution (violating items are a tiny fraction of total traffic).

Adversarial manipulations in images and product structures.

Solution Overview

The pipeline combines three complementary models:

Brand classification model (image‑level).

Brand logo detection model (object detection).

Brand image‑variation recognition model (captures collage/stacking patterns).

A large‑scale labeled dataset was built to train and evaluate these models.

Dataset Construction

The dataset covers 212 brands with 180,605 images (average 852 images per brand) and >97% labeling accuracy. It includes:

Brand feature definitions (e.g., key visual cues for GUCCI).

Classification data : all images of a brand are grouped into a single class, regardless of category or style.

Detection data : each image is annotated with a rectangular box around a single brand logo to simplify labeling.

Data sources: product catalog, penalty records, user‑generated comment images, and targeted keyword searches for important styles.

Brand Classification Model

A binary classifier was first validated on Nike (10k positive, 10k negative samples). Higher confidence thresholds increase precision at the cost of recall. For production, each brand is treated as one class and all non‑brand items as a second class, enabling low‑cost inference (~4,000 QPS on a single G41 server).

Evaluation and Optimization

Three test sets were used:

Closed‑set test (212 brands) : average recall 58.6% (base) → 60.0% after optimization.

Open‑set test (5 M random samples) : recall proportion 0.32% at confidence > 0.5.

Business test (historical violation data) : recall without logo 25.5% → 45.8%; with logo 51.1% → 62.5% after optimization.

Optimization steps:

Balance negative samples per brand.

Emphasize hard‑negative examples.

Iteratively focus on brands with low precision.

Brand Logo Detection Model

A one‑stage YOLOv3 detector was chosen for speed and small‑object performance. The base model achieved 78.06% mAP on the validation set.

Enhancements

Data augmentation : random translation, soft labels, MixUp, Mosaic (from YOLOv4/v5).

Loss redesign : focal loss for classification (positive weight 0.7) and DIoU loss for bounding‑box regression.

Network redesign : replace Darknet‑53 with a CSPNet‑based backbone (SPP + Focus modules), reducing model size to 44.5 MB, inference time to 24 ms, and raising mAP to 89.89%.

Model Fusion

Combining image classification, text classification, and logo detection yields complementary strengths. Fusion improves average recall and online recall proportion by >10 percentage points compared with any single model.

Future Work

Expand brand coverage to include more categories.

Strengthen adversarial robustness through targeted training.

Explore semi‑supervised and unsupervised learning to further improve recognition performance.