Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

XAI-Based Detection of Adversarial Attacks on Deepfake Detectors

Authors: Ben Pinhasov, Raz Lapid, Rony Ohayon, Moshe Sipper, Yehudit Aperstein

TMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted experiments to assess the performance of the deepfake detectors based on Xception Net and Efficient Net B4ST, following the methods described by Chollet (2017) and Bonettini et al. (2021). We used the Face Forensics++ (FF++) dataset (Rossler et al., 2019) to assess the effectiveness of our proposed face-forgery detection method.
Researcher Affiliation Collaboration Ben Pinhasov EMAIL Afeka, The Academic College of Engineering in Tel Aviv, Raz Lapid EMAIL Ben-Gurion University of the Negev, Beer-Sheva, 8410501, Israel & Deep Keep, Tel-Aviv, Israel, Rony Ohayon EMAIL Deep Keep, Tel-Aviv, Israel
Pseudocode No The paper only describes the methodology and algorithms in prose and through a flowchart (Figure 3), without specific pseudocode or algorithm blocks.
Open Source Code Yes Code available at https://github.com/razla/XAI-Based-Detection-of-Adversarial-Attacks-on-Deepfake-Detectors
Open Datasets Yes We used the Face Forensics++ (FF++) dataset (Rossler et al., 2019) to assess the effectiveness of our proposed face-forgery detection method.
Dataset Splits Yes For our experiments, we randomly picked 60 videos from this dataset, with 50 allocated for training (including validation) and 10 for testing, resulting in 20,000 images for training and 5,000 images for testing, per each model and XAI configuration.
Hardware Specification Yes The experimental setup included a high-performance computing environment with an Intel Xeon Gold 6136 @ 3Ghz CPU, NVIDIA GRID P40-24Q, and 64GB of RAM.
Software Dependencies Yes The software environment consisted of Windows 10, Python 3.9, Pytorch 2.0.1, and the Captum library for implementing XAI techniques (Kokhlikyan et al., 2019).
Experiment Setup Yes Detect-Res Net50 model was trained with a learning rate of 0.001, a minibatch size of 16, and 100 iterations. We used the cross-entropy loss function, and the Adam optimizer (Kingma & Ba, 2014) with hyperparameters ϵ = 1e 8, β1 = 0.9, and β2 = 0.999.