FatimahEmadEldin/genai-manipulation-detection-interior

GenAI Manipulation Detection Dataset - Interior Design Images

📋 Dataset Description

This dataset contains 1000 paired images (real + manipulated) for training and evaluating GenAI manipulation detection models. Created for the MenaML Winter School 2026 Hackathon.

Dataset Summary

• Total Images: 1000 pairs (2000 total images)
• Image Size: 512x512
• Format: JPEG
• Source: Pinterest Interior Design Images (Kaggle)
• License: MIT

🎯 Challenge Context

This dataset was created for Track B: Real Estate & Commercial Integrity of the MenaML Winter School 2026 GenAI Detection Challenge.

The challenge focuses on detecting:

• ✅ Virtual staging (furniture replacement)
• ✅ Texture smoothing (wall/surface manipulation)
• ✅ Compression artifacts
• ✅ Splicing and copy-move forgery
• ✅ Physical impossibilities (shadow/reflection mismatches)

📂 Dataset Structure

dataset/
├── data/
│   ├── real/              # Original unmanipulated images
│   │   ├── real_000000.jpg
│   │   ├── real_000001.jpg
│   │   └── ...
│   └── fake/              # Manipulated images
│       ├── fake_000000.jpg
│       ├── fake_000001.jpg
│       └── ...
├── annotations.csv               # ⭐ Main annotations (2 rows per pair)
├── detailed_annotations.json     # Paired format annotations
├── metadata.json                 # Dataset statistics
└── README.md                     # This file

📊 Annotations Format

CSV Annotations (annotations.csv)

Each image pair creates 2 rows - one for real, one for fake:

file_name,image_path,label,is_manipulated,manipulation_category,manipulation_technique,manipulation_description,image_id,pair_id
real_000000.jpg,data/real/real_000000.jpg,real,0,none,none,Authentic unmanipulated image,000000,000000
fake_000000.jpg,data/fake/fake_000000.jpg,fake,1,smoothness_anomaly,bilateral_filter,Unnatural smoothness in walls/surfaces,000000,000000
real_000001.jpg,data/real/real_000001.jpg,real,0,none,none,Authentic unmanipulated image,000001,000001
fake_000001.jpg,data/fake/fake_000001.jpg,fake,1,compression_artifact,double_jpeg,Double JPEG compression,000001,000001

Columns:

• file_name: Image filename
• image_path: Relative path to image
• label: "real" or "fake"
• is_manipulated: 0 (real) or 1 (fake)
• manipulation_category: Category (or "none" for real images)
  • smoothness_anomaly
  • compression_artifact
  • frequency_manipulation
  • splicing
  • physical_impossibility
• manipulation_technique: Specific technique (or "none" for real images)
• manipulation_description: Human-readable description
• image_id: Unique ID for this image
• pair_id: ID linking real and fake pairs (same for both images in a pair)

JSON Annotations (detailed_annotations.json)

Paired format for easier processing:

[
  {
    "pair_id": "000000",
    "real_image": {
      "filename": "real_000000.jpg",
      "path": "data/real/real_000000.jpg",
      "label": "real"
    },
    "fake_image": {
      "filename": "fake_000000.jpg",
      "path": "data/fake/fake_000000.jpg",
      "label": "fake",
      "manipulation": {
        "category": "smoothness_anomaly",
        "technique": "bilateral_filter",
        "description": "Unnatural smoothness in walls/surfaces"
      }
    }
  }
]

📊 Manipulation Categories

• compression_artifact: 191 images (19.1%)
• smoothness_anomaly: 206 images (20.6%)
• physical_impossibility: 192 images (19.2%)
• frequency_manipulation: 209 images (20.9%)
• splicing: 202 images (20.2%)

Detailed Technique Breakdown

• compression_mismatch: 100 (10.0%)
• bilateral_filter: 108 (10.8%)
• texture_removal: 98 (9.8%)
• reflection_inconsistency: 94 (9.4%)
• upscaling: 74 (7.4%)
• copy_move: 94 (9.4%)
• frequency_injection: 70 (7.0%)
• object_insertion: 108 (10.8%)
• grid_artifact: 65 (6.5%)
• shadow_mismatch: 98 (9.8%)
• double_jpeg: 91 (9.1%)

🔧 Manipulation Techniques Explained

1️⃣ Smoothness Anomaly

• bilateral_filter: Aggressive bilateral filtering creating unnatural smoothness
• texture_removal: Edge-preserving filter that removes texture detail
• Detection: Texture analysis, high-frequency loss detection

2️⃣ Compression Artifact

• double_jpeg: Two rounds of JPEG compression with different quality levels
• compression_mismatch: Regions with different compression quality
• Detection: DCT coefficient analysis, block artifact detection

3️⃣ Frequency Manipulation

• upscaling: Downscale→upscale creating bicubic interpolation signatures
• frequency_injection: GAN-like ring patterns in frequency domain
• grid_artifact: 8×8 grid patterns typical of GAN outputs
• Detection: FFT analysis, power spectral density

4️⃣ Splicing

• copy_move: Copy region and paste elsewhere with compression mismatch
• object_insertion: Insert objects with different compression characteristics
• Detection: SIFT/ORB feature matching, noise inconsistency analysis

5️⃣ Physical Impossibility

• shadow_mismatch: Inconsistent shadow directions
• reflection_inconsistency: Reflections not matching room layout
• Detection: VLM reasoning, physics-based validation

💻 Usage

Load All Data

import pandas as pd
from PIL import Image
# Load annotations
df = pd.read_csv("hf://datasets/FatimahEmadEldin/genai-manipulation-detection-interior/annotations.csv")
print(f"Total images: {len(df)}")
print(f"Real images: {len(df[df['label'] == 'real'])}")
print(f"Fake images: {len(df[df['label'] == 'fake'])}")
# Show manipulation distribution (fake images only)
fakes = df[df['label'] == 'fake']
print("\nManipulation categories:")
print(fakes['manipulation_category'].value_counts())

Filter by Manipulation Type

# Get only images with specific manipulation
smoothness_fakes = df[df['manipulation_category'] == 'smoothness_anomaly']
print(f"Smoothness anomalies: {len(smoothness_fakes)}")
# Get specific technique
bilateral_images = df[df['manipulation_technique'] == 'bilateral_filter']
print(f"Bilateral filter: {len(bilateral_images)}")
# Get all compression artifacts
compression = df[df['manipulation_category'] == 'compression_artifact']
techniques = compression['manipulation_technique'].value_counts()
print(techniques)

Load Image Pairs

# Get a pair by pair_id
pair_id = "000000"
pair = df[df['pair_id'] == pair_id]
real_row = pair[pair['label'] == 'real'].iloc[0]
fake_row = pair[pair['label'] == 'fake'].iloc[0]
real_img = Image.open(f"hf://datasets/FatimahEmadEldin/genai-manipulation-detection-interior/{real_row['image_path']}")
fake_img = Image.open(f"hf://datasets/FatimahEmadEldin/genai-manipulation-detection-interior/{fake_row['image_path']}")
print(f"Real: {real_row['file_name']}")
print(f"Fake: {fake_row['file_name']}")
print(f"Manipulation: {fake_row['manipulation_technique']}")
print(f"Description: {fake_row['manipulation_description']}")

Training with Class Labels

# Create dataset with manipulation classes
fakes_only = df[df['label'] == 'fake'].copy()
# Map techniques to numeric labels
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
fakes_only['class_label'] = le.fit_transform(fakes_only['manipulation_technique'])
print("Class mapping:")
for i, tech in enumerate(le.classes_):
    print(f"  {i}: {tech}")
# Use for training
from torch.utils.data import Dataset
class ManipulationDataset(Dataset):
    def __init__(self, dataframe, transform=None):
        self.df = dataframe.reset_index(drop=True)
        self.transform = transform
        
    def __len__(self):
        return len(self.df)
    
    def __getitem__(self, idx):
        row = self.df.iloc[idx]
        img = Image.open(row['image_path'])
        
        if self.transform:
            img = self.transform(img)
        
        return {
            'image': img,
            'label': row['class_label'],
            'category': row['manipulation_category'],
            'technique': row['manipulation_technique']
        }

Binary Classification (Real vs Fake)

# Simple real vs fake
df['binary_label'] = df['is_manipulated']  # 0 = real, 1 = fake
# Or use the label column
df['binary_label'] = (df['label'] == 'fake').astype(int)
# Split by label
real_images = df[df['binary_label'] == 0]
fake_images = df[df['binary_label'] == 1]

Multi-task Learning

# Train on both binary and multiclass
class MultiTaskDataset(Dataset):
    def __init__(self, dataframe, technique_to_idx, transform=None):
        self.df = dataframe.reset_index(drop=True)
        self.technique_to_idx = technique_to_idx
        self.transform = transform
        
    def __getitem__(self, idx):
        row = self.df.iloc[idx]
        img = Image.open(row['image_path'])
        
        if self.transform:
            img = self.transform(img)
        
        # Binary label
        binary_label = row['is_manipulated']
        
        # Multiclass label (only for fake images)
        if binary_label == 1:
            multiclass_label = self.technique_to_idx[row['manipulation_technique']]
        else:
            multiclass_label = -1  # Ignore for real images
        
        return {
            'image': img,
            'binary_label': binary_label,
            'multiclass_label': multiclass_label,
            'category': row['manipulation_category']
        }

📝 Annotations Format

The dataset includes two annotation files for different use cases:

1. annotations.csv - Flat Format

Contains 2 entries per pair (one for real, one for fake):

Column	Description	Example Values
file_name	Image filename	real_000000.jpg, fake_000000.jpg
image_path	Relative path	data/real/real_000000.jpg
label	Real or fake	real, fake
is_manipulated	Binary flag	0 (real), 1 (fake)
manipulation_category	Category	smoothness_anomaly, compression_artifact, etc.
manipulation_technique	Specific technique	bilateral_filter, double_jpeg, etc.
manipulation_description	Human description	"Unnatural smoothness in walls..."
image_id	Unique image ID	000000
pair_id	Links real/fake pairs	000000 (same for both in pair)

Use this for:

• Quick filtering by label or technique
• Training binary classifiers
• Training multiclass classifiers
• Statistical analysis

2. detailed_annotations.json - Paired Format

Contains paired structure (1 entry per pair): Use this for:

• Contrastive learning
• Side-by-side comparison
• Paired image processing

Manipulation Categories

1. smoothness_anomaly (20%)

   • bilateral_filter: Aggressive bilateral filtering
   • texture_removal: Edge-preserving texture removal

2. compression_artifact (20%)

   • double_jpeg: Multiple JPEG compression
   • compression_mismatch: Regional quality differences

3. frequency_manipulation (20%)

   • upscaling: Bicubic upscaling artifacts
   • frequency_injection: GAN-like frequency patterns
   • grid_artifact: 8×8 grid patterns

4. splicing (20%)

   • copy_move: Copy-paste forgery
   • object_insertion: Object insertion with mismatches

5. physical_impossibility (20%)

   • shadow_mismatch: Inconsistent shadow directions
   • reflection_inconsistency: Impossible reflections

🎓 Citation

If you use this dataset in your research or hackathon submission, please cite:

@dataset{genai_manipulation_interior_2026,
  title={GenAI Manipulation Detection Dataset - Interior Design},
  author={MenaML Winter School 2026 - Team ArtifactDetect},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/FatimahEmadEldin/genai-manipulation-detection-interior}}
}

📜 License

This dataset is released under the MIT License. The source images are from the Pinterest Interior Design Images dataset on Kaggle, used under MIT license.

🏆 Hackathon Information

• Event: MenaML Winter School 2026
• Challenge: Detecting GenAI & Sophisticated Manipulation in Public Media
• Track: B - Real Estate & Commercial Integrity
• Deadline: January 28, 2026

🙏 Acknowledgments

• Source dataset: Pinterest Interior Design Images by Galina KG
• Challenge organizers: MenaML Winter School 2026
• Tools: OpenCV, scikit-image, NumPy, Pandas