# Response to R1

We thank the **Reviewer K6rp** for the thoughtful review and the vote to accept the paper. Below we provide a response to key questions/comments:

## Application to other deepfake techniques

> This paper aim for GAN-generated deepfake images and if it can be applied to distinguish other local deepfake forgery techiques (Deepfakes, Face2Face, FaceSwap and NeuralTextures, etc).

As per prior work [1], the frequency artifacts observed in GAN-generated images are a result of the transpose convolution operation used in the GAN generators. Since other deepfake forgery techniques are not guaranteed to employ this operation, it is unclear but an interesting question as to whether these frequency artifacts are present and redundant in images produced via other techniques. We note that only the “Deepfakes” and “NeuralTextures” techniques mentioned above use generators with transpose convolutional layers (the other two are non-neural). As a preliminary result, we present below results from training and evaluating the natural performance of our best-performing ensemble, D3-S(4) against the AT baseline on the “Deepfakes” and “NeuralTextures” subsets of the FaceForensics++ dataset [2].

| Dataset | AT (1) | D3-S (4) |
| ----------|--------|-----------|
|Deepfakes| 98.8 | 90.6 |
|NeuralTextures | 94.6 | 80.8 |

The respectable natural accuracy of D3-S(4) compared to the full-spectrum AT baseline suggests that redundant frequency features likely also apply to deepfakes generated using other neural techniques, so long as transpose convolution is involved in the generation upsampling. Note that the reduced accuracy of AT for these datasets (as compared to GANs, where the accuracies were >99.9%) suggests that these images do have fewer frequency artifacts. This translates to reduced performance for D3-S(4) as well, since each individual model has access to a subset of frequencies. We will include further numbers on adversarial accuracies in the final draft.

### References in this response
- [1] Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. 2020. Leveraging frequency analysis for deep fake image recognition. In Proc. of ICML.
- [2] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. 2019. Faceforensics++: Learning to Detect Manipulated
Facial Images. In Proc. of ICCV.