aics-project

Synthetic vs. genuine image detection

An exploration of ML for the detection of AI-generated images which imitate photographic realism based on Bird & Lotfi (2024) and the CIFAKE (Bird, 2023) dataset.

Required datasets, to be placed in code/:
CIFAKE https://www.kaggle.com/datasets/birdy654/cifake-real-and-ai-generated-synthetic-images/data
CIFAR100: https://www.kaggle.com/datasets/fedesoriano/cifar100/data
ai-generated-images-vs-real-images/test: https://www.kaggle.com/datasets/tristanzhang32/ai-generated-images-vs-real-images (only the test dir!, but include the 'ai-generated-images-vs-real-images' folder for path structure)

MLT-GPU path: /srv/data/gussucju/aics-project
However, I had trouble uploading the datasets & found different info re whether to supply or just link these. Since my main datasets (CIFAKE, CIFAR100) are pretty small & training is quick, I opted to give download instructions for these. AGIRI is the only larger dataset, let me know if I should supply these in some other way.

---

Quickstart:

• Download the required datasets as noted above, place in code/
• (pretrained model files are stored in models/; for training of new models follow the next two bulletpoints in order)
• Go to code/CIFAKE_classification.ipynb or code/func_difficulty.ipynb for training/testing on CIFAKE
• Go to transfer_cifar100.ipynb or transfer_mixed.ipynb to load & apply models on CIFAR100/AGIRI

---

Repos breakdown:

code/: all code should go in this folder. For coding we recommend using Jupyter notebooks. All steps required to run the code should be well documented, as well as individual parts of the code should be explained with comments. One should be able to run the code and replicate the process.

Notebooks

CIFAKE_classification.ipynb

• Training & testing of the base model and model modified with SRM channel attention

func_difficulty

• Exploration of CIFAKE task difficulty; examine training progress of first epoch after X batches

transfer_cifar100.ipynb

• Testing the transfer performance of base, mini, and attention model on cifar100 (unseen, real images)
• Viewing performance per super- and atomic class

transfer_mixed.ipynb

• Testing the transfer performance of base, mini, and attention model on dataset of ai/real imgs from various origins/styles (Zhang)

Scripts

classes.py

OOP classes for...

• CI_loader: data batching, transformation, encoding; used for CIFAKE & CIFAR100
• CIFAKE_CNN: Model architecture based on course tutorial notebook supervised/cifar10_tutorial.ipynb (https://canvas.gu.se/files/9275057/), modified for binary classification
• SRMLayer: Style-recalibration module channel attention layer (https://blog.paperspace.com/srm-channel-attention/, detailed sources in code+paper)

train.py

• Script to train models on CIFAKE
• cmd line args: batch size (bs; 32), epochs (ep; 5), modelfile (mf; only save model if passed; format: '../models/{args.modelfile}.pth'
  Example call, from code/:

$ python3 code/train.py -ep 1 -mf dummymodel

test.py

• Script to test & evaluate model on CIFAKE, CIFAR100 test data, includes eval functions used in notebooks
• cmd line args: modelfile (mf; model to test from '../models/{args.modelfile}.pth', default: base_model), decision_threshold (thr; 0.5)
  Example call, from code/:

$ python3 code/test.py -thr 0.4

config.json

• simple json file to keep track of some filenames; I.e. CIFAKE dir, base/attention/mini model paths

---

data/: use this folder for datasets that you have created. Those that can be downloaded from elsewhere can be referenced in README.md and files that are created by your code should be in your folder on MLTGPU. If you need to upload large files to Github, you can use their Git Large File Storage.

• (CIFAKE should be placed in here)
• (CIFAR100 should be placed in here)
• (ai-generated-images-vs-real-images (AGIRI) should be placed in here, but we only need the /test folder therein)

---

library/: a place for any supporting files that you have used, for example required system configurations and other supporting files that cannot be obtained from open repositories.

• (not used)

---

models/: place to store trained models for simple loading

• shared params: batch size 32, learn rate 0.001, momentum 0.9

base_model.pth

• base CIFAKE_CNN, trained on CIFAKE train data, epochs 10

mini_model.pth

• base CIFAKE_CNN, trained on CIFAKE train data for 1 epoch to examine task difficulty

attn_model.pth

• CIFAKE_CNN with SRM channel attention layer before 2nd conv2d layer, trained on CIFAKE train data (same params as base)

---

notes/: To show the work that you have done (especially important if you get a negative result) you should keep a lab log as you work. This will also allow you to write the report faster. Use this folder with markdown text. files as a wiki to document your experiment and take notes as you go along. These files can also be edited and viewed on Github.

lognotes.md

• work progress diary (messy, personal use)

project_pitch_og.md

• copy of original project pitch (LT2318_AICS_project_pitch.pdf) as markdown file

project_outline.md

• updated, short overview of project steps

---

paper/: a place for your course report and presentation

LT2318_AICS_project_pitch.pdf

• original project pitch; proposed project steps
• updated on canvas discussion thread

aics_report.pdf

• project paper

aics_ppp.pptx

• presentation slides (draft)