MesoNet: the 2018 mesoscopic deepfake detector and how it scores today

This page sits in the published results section of the deepfake detection hub. It reads one detector's row across time; the Xception detector page does the same for the network that replaced MesoNet as the usual baseline.

01What is MesoNet?

MesoNet is the name the paper gives to two architectures, Meso-4 and MesoInception-4, written Meso4 and MesoIncep in later benchmark tables (arXiv 1809.00888). Meso-4 is four blocks of convolution and pooling followed by a dense layer with one hidden layer, and has 27,977 trainable parameters. MesoInception-4 replaces the first two convolutional layers with a variant of the inception module that uses 3 by 3 dilated convolutions in place of 5 by 5 ones, and has 28,615 parameters. Both take 256 by 256 face crops.

The design argument is in the name. The authors write that microscopic analyses based on image noise "cannot be applied in a compressed video context where the image noise is strongly degraded", and that at the semantic level the human eye struggles with forged faces. A network with few layers, they argue, settles at an intermediate level of representation that survives H.264 compression. The code was released on GitHub with the paper, which is one reason MesoNet became a fixture in later comparisons.

02What did the 2018 paper report?

The paper built its own Deepfake set because, in the authors' words, "no dataset gathers videos generated by the Deepfake technique". They collected 175 forged video rushes from online platforms, at a minimum resolution of 854 by 480 pixels and with varied H.264 compression, and extracted faces into 5,111 forged and 7,250 real training images and 2,889 forged and 4,259 real test images. For Face2Face they used the original FaceForensics set at its three compression levels (0, 23 and 40), with 300 videos for training and the 150 test videos for evaluation: 4,500 forged and 4,500 real training faces, 3,000 and 3,000 for testing.

Scored frame by frame, Meso-4 reached 0.891 and MesoInception-4 0.917 on the Deepfake set (Table 3). On Face2Face, Meso-4 scored 0.946, 0.924 and 0.832 at compression 0, 23 and 40, and MesoInception-4 0.968, 0.934 and 0.813 (Table 4). The headline figures come from averaging the per-frame predictions over a whole video: 0.969 for Meso-4 and 0.984 for MesoInception-4 on Deepfake, and 0.953 for both on Face2Face at compression 23 (Table 5). The abstract's "more than 98% for Deepfake and 95% for Face2Face" is this video-level number, on the authors' own test split.

03How does MesoNet score in DeepfakeBench?

DeepfakeBench (NeurIPS 2023 Datasets and Benchmarks) retrained both networks under one protocol, with every detector trained on FaceForensics++ c23 and scored by frame-level AUC on six within-domain and eight cross-domain test sets (arXiv 2307.01426). The table below reads the two MesoNet rows from its Table 3; the full grid with all fifteen detectors is on the frame-level results table.

Meso4 holds the lowest within-domain average of the fifteen detectors, 0.6097, and the lowest cross-domain average, 0.6551. Its 0.5481 on Google DFD is the lowest single cross-domain cell in the table. MesoIncep does better everywhere, and both networks score above 0.90 on DeeperForensics-1.0, where Xception scores 0.8341. The paper groups both under "Naive" detectors, meaning a CNN that "directly distinguish[es] deepfake content from authentic data" with no hand-designed cue.

04Why does a 2018 network score near chance on a 2023 table?

The two sets of numbers measure different things. The 2018 paper trained and tested on the same kind of fake, scored per video, and reported accuracy. DeepfakeBench trains on FaceForensics++ c23, scores per frame, and reports AUC, so a single misread frame counts in full and there is no averaging to smooth it. The authors of MesoNet themselves showed how much aggregation adds: Meso-4 went from 0.891 per frame to 0.969 per video on their Deepfake set.

Capacity is the other factor. A network with 27,977 parameters was built to find one kind of blur boundary in two kinds of fake. FaceForensics++ c23 holds four manipulation methods, and the benchmark's cross-domain sets add generators the network has never seen. The DeepfakeBench codebase page describes the protocol; its Table 3 shows that Meso4's weakest within-domain cell is NeuralTextures at 0.5701, a method that edits only the mouth region and did not exist when MesoNet was designed. Detectors that watch mouth motion over time, covered on the lip-sync deepfakes page, were built for exactly that gap.

05MesoNet in other published comparisons

MesoNet appears as a baseline in several of the papers the site tracks, and its figures move with the protocol each time. In the FaceForensics++ paper (ICCV 2019), trained on all four manipulations, MesoNet reached a binary accuracy of 95.23 percent on raw video, 83.10 percent at HQ (c23) and 70.47 percent at LQ (c40), against 99.26, 95.73 and 81.00 for XceptionNet (arXiv 1901.08971, Table 2). In WildDeepfake (ACM Multimedia 2020), MesoNet-4 scored 97.04 percent image-level accuracy on the FaceForensics++ Deepfake HQ split and 64.47 percent on fakes collected from the internet; MesoNet-inception scored 97.16 and 66.03 (arXiv 2101.01456, Table 3).

Read together, the three comparisons agree on the shape: MesoNet is competitive on the fake it was trained on, loses more than its deeper rivals under compression, and falls toward chance on fakes from another source. No single benchmark score predicts how a detector performs on unseen manipulations, and MesoNet is the clearest case in the table.

06How should a MesoNet figure be read?

Any MesoNet number needs four conditions beside it: which variant (Meso-4 or MesoInception-4), which training set, whether the score is per frame or per video, and whether it is accuracy or AUC. The 0.984 of the 2018 paper and the 0.6077 of the 2023 table are both correct, and neither describes the other's setting. A benchmark that retrains every detector under one protocol, as DeepfakeBench does, is the only place the two architectures can be compared with the rest of the field; a paper's own table shows what its authors could achieve on their own data.

The same reading discipline applies on the recognition side of this site, where a face recognition test is defined by its pairs, threshold policy and metric before any score is quoted.