Semantic Calibration (2025)

This repository contains the official implementation of our paper.

🔗 Links

• 📝 Paper: [2025 PDF Link]
• 🤗 Hugging Face: https://huggingface.co/BeyondDeepFakeDetection
• 📊 Datasets: https://huggingface.co/BeyondDeepFakeDetection
• 📦 Fine-tuned Models: https://huggingface.co/BeyondDeepFakeDetection
• 📁 Code: https://github.com/BeyondDeepFakeDetection/Beyond-Deep-Fake-Detection

---

🧠 Overview

Semantic Calibration addresses a key limitation in current deepfake detection systems: the overreliance on low-level visual artifacts that may disappear as generative models improve. Instead of merely asking "Is this media fake?", our work shifts the focus to a more meaningful question: "Is the semantic content deceptive?"

We formally define the notion of deception in online media and introduce semantic calibration — a framework that detects media samples which distort semantic expectations, even when they are synthetically flawless. Our pipeline operates by:

• Captioning media content into text using image-to-text and audio-to-text models
• Computing acceptance probabilities for semantic consistency using fine-tuned large language models (LLMs)
• Rejection sampling over the semantic space to detect content that is surprising or overrepresented under a calibrated distribution

We release:

• Code to reproduce all experiments from the paper.
• Pretrained GPT-2 models fine-tuned on general and real distributions.
• Semantically distorted datasets simulating distribution shifts.
• Evaluation script for kl-divergence and deception reduction metrics.

---

Visualization of our filtering mechanism

The live figure below shows how our filtering mechanism operates in an online setting. The yellow bars represent the general distribution of label IDs, which is the original distribution from which we sample. The dark blue bars represent the real distribution, which is our target. The light blue bars show the filtered distribution, which is updated continuously and should closely match the real distribution. In the top right corner, a legend displays the number of accepted samples, and just next to it, the current deception reduction level is shown in real time.

Repository Structure

creating_datasets/

1. distort.py
   Generates various splits of the real data distribution from the original general distribution. You can specify a target KL-divergence and adjust the target_frequencies list to obtain a subset that meets the desired divergence.

2. Label Mapping JSONs
   JSON files used to map fine-grained labels into custom coarse-grained labels.

---

scripts/

1. Florence_captioning.py
   Contains the exact configuration used to generate image captions for our image datasets using the Florence model.

2. finetune_gpt2.py
   Script used to fine-tune 40 GPT-2 models with our experimental setup.

3. qwen_audio_captioning.py
   Generates audio captions using Qwen-Audio, aligned with the methodology in our experiments.

4. semantic_calibration_evaluation.py
   Main evaluation pipeline script. Handles rejection sampling, computes performance metrics (deception reduction), and assumes precomputed log-probabilities per token (via tokens_sequence_probs.py).

5. tokens_sequence_probs.py
   Computes per-token log-probabilities and overall sequence likelihoods using fine-tuned GPT-2 models. Used as input for the evaluation script.
   
   
   

📦 Installation

We recommend using a virtual environment:

git clone https://github.com/BeyondDeepFakeDetection/.git
cd Beyond-Deep-Fake-Detection
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

✅ Quick Start

We will navigate through an example of running the full pipeline and reproducing results from the paper.

Computing log-probs of texts

1. Install dependencies:

   pip install transformers torch pandas huggingface_hub safetensors tqdm datasets seaborn matplotlib numpy 

2. Set the following variables in token_sequence_probs.py:

   • hf_token: Your HuggingFace API token.
   • model_name: The name of the model to use (e.g., "CIFAR-10_general, CIFAR-10_moderate").
   • save_name: Path (without .csv) where token-prob results will be saved.
   • dataset: The HuggingFace dataset name (e.g., "CIFAR-10").
   • text_column: The column in the dataset containing the text (default: "text"). This script computes log probabilities for each word in a dataset of texts using a HuggingFace language model, and saves the results to a CSV file.

3. Run the script twice, once for the general and one for the real model with the split of choice (baseline|mild|moderate|severe):

   python token_sequence_probs.py

Output token_sequence_probs.py

The script saves a CSV file at <save_name>.csv with columns:

• text_id: Index of the text in the dataset.
• total_log_prob: Sum of log probabilities for the text.
• avg_log_prob: Average log probability per word.
• word_probabilities: List of (word, log_prob) pairs. Now having two such CSV files we can compare probability ratios between the two models.

Running the rejection sampling and evaluating performance

4. Set the following variables in semantic_calibration_evaluation.py

   • "label_column_name": The column in teh dataset containing the label (fine_label, coarse_label, label, etc.)
   • "text_column_name" : The column in the dataset containing the text (default: "text").
   • "split_name": choice of split name (baseline/mild/moderate/severe)
   • "dataset_name" : The HuggingFace dataset name (e.g., "CIFAR-10")
   • "save_path" : Set the name o the output path for the Label Distribution Comparison plot.
   • "real_probs_csv" : Enter the output path of the real-model probabilities (computed in step 3)
   • "general_probs_csv" : Enter the output path of the general-model probabilities (computed in step 3)

5. Run the script:

   python semantic_calibration_evaluation.py

Output semantic_calibration_evaluation.py

Printed you will find the results of the experiment, namely the KL-divergence between the general and real distribution, and the filtered and real distribution. The goal is to minimize the latter. A matplotlib chart of the normalized frequencies of the labels will also be plotted.