Task Overview

Dataset Overview

To address this challenge, we construct the DPATeX dataset. We begin by manually collecting 257 perturbation cores (PCs) from the literature, each expressing a core adversarial mechanism, and storing them in LaTeX format. These curated PCs serve as seeds to prompt a large language model (LLM) to generate 256 additional perturbation cores. The generation process adheres to strict control criteria: each generated core must belong to the same mathematical family as its seed, remain distinct, and preserve its adversarial effectiveness. The resulting DPATeX dataset comprises a total of 513 perturbation cores, categorized into seven mathematical classes: Gradient, Norm, Hybrid, Machine Learning (ML), Geometric, Bayesian Optimization, and Gradient-Free. Each sample is labeled and formatted in LaTeX, providing a structured resource for advancing AI-based defense research.

Each sample in the dataset consists of either a mathematical formula or an algorithm. Each DPATeX sample is stored in a LaTeX-based file, organised into four main sections:

Input: Defines the mathematical input variables required for the perturbation function.

Output: Specifies the perturbed mathematical output after applying the perturbation function.

Formula: Contains the detailed mathematical formulation written in LaTeX. This section defines the precise operations, transformations, and perturbation functions used to generate the output.

Explanation: Provides a descriptive clarification of the formula, offering additional insights into the perturbation function and its role in data poisoning.

Task Description

This competition task is to distinguish machine generated PCs from the seed PCs, which are developed by humans. The full dataset consists of the seed PC dataset DPATeX, machine generated DPAGenTeX, and the merged dataset. The statistical information of the dataset used for this competition is given as:

Dataset	Samples/Files in Two classes 0 is seed, 1 is machine generated
Train	410
Test	103

The name of file identifies the label of the seed (human complied) PC versus machine generated PC as "file name_0.tex" and "file name_1.tex", respectively.

Results Submission

You need to submit a CSV file that contains the test sample predictions. The expected result submission should be a CSV file with 2 columns with the first column as the file name and second column as the class label.

Citation of this Dataset

Task Overview

Dataset Overview

To facilitate this task, we provide a dataset comprising Function Call Graphs (FCGs) extracted from real-world IoT malware samples targeting ARM CPU architectures. Each sample captures the dynamic interaction between malware functions, represented as:

• .data file: Contains FCG representations where:
• Each line corresponds to one malware sample
• Edges are defined as 2-tuples of function names (caller#callee), separated by spaces
• Example: funcA#funcB funcC#funcD represents a graph with two edges: funcA → funcB and funcC → funcD
• .label file: Provides supervised labels for training, with one class label per line (aligned with the .data file entries).

The dataset is partitioned into:

Dataset	Samples
Train	1,397
Test	1,397

Task Description

Participants must:

1. Train a classifier using train.data and train.label to learn malware categories from FCGs.
2. Predict labels for unseen samples in test.data, generating one label per line.
3. Submit results in a text file mirroring train.label format.

Evaluation Focus:

• Accuracy in classifying malware families
• Scalability to large FCGs
• Robustness to obfuscation techniques

Submission Guidelines

Submit a .CSV file with one predicted label per line for test.data, e.g.:

mirai gafgyt gafgyt rootnik mobidash tsunami ddostf

Technical Hints

1. Basic Approach: Preprocess edges (replace # with spaces) to extract caller-callee pairs.
2. Advanced Methods: Leverage graph topology (e.g., node degrees, connectivity patterns) or graph neural networks (GNNs) for higher accuracy.

Citation

Task Overview

Task Description

The objective of this competition task is to develop a robust multimodal system that can accurately predict answers by processing both textual questions and their corresponding photographed table images. The challenge spans multiple STEM disciplines, including mathematics, physics, chemistry, biology, among others, requiring models to demonstrate effective cross-modal reasoning under real-world conditions. While the dataset uses Chinese tables, the competition task itself is language-agnostic as it evaluates the model's reasoning capabilities rather than language proficiency.

Dataset Overview

The PCTR dataset consists of 13,298 training samples and 1,000 test samples. The training data contains real-world noise and some annotation errors, while the test set has been meticulously verified by experts to ensure accuracy.

All data is organized in JSON format with accompanying image files. The training set includes train/train.json and a corresponding train/images directory, while the test set follows the same structure with test/test.json and test/images. Each JSON entry contains: a unique question ID, the file path to the associated photographed table image, a question grounded in the image content, a step-by-step annotated solution, and the final answer represented as a string. Below gives an example of JSON entry and corresponding image for training.

{"id": "4", "image": "images/train/1622044524477847261674147110912_0.jpg", "question": "初一和初二在90≤x≤100分数段的总人数是多少?", "solution": "12+15", "answer": "27"}

For testing, test.json gives the question and file path to corresponding image file.

{"id": "1", "image": "images/test/dfdb38e3500edd96d45b3398ae8b0e65.jpg", "question": "如果从周六之后开始之后到下周六利润呈等差数列，那下周三的利润是多少元？"}

Results Submission

For performance evaluation, participants should submit their model's predictions in a CSV file, including two columns:

• id: The unique ID of the question in the test set.
• model_answer: The model's predicted answer formatted as a string without units (e.g., "1538" instead of "1538 人". For yes/no questions, the answer should be either "是" or "否" in Chinese.

Your submitted CSV answer should be formatted as the example given below:

id, model_answer

"1", "**"

"2", "**"

Citation of this Dataset