A deepfake detector for faces was tested in a lab, not in an election

experiment · PloS one · la publicación, 6 abr 2026 · gratis

Herramienta de laboratorio detecta caras falsas con 99% de acierto, pero no se probó en una elección real ni en su comunidad

Versión breve · la versión detallada sigue, unos 5 min

Pregunte a weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.

El estudio, de un vistazo
Quiénes: Rostros reales y falsos, incluidos 500 rostros reales de personas de Bangladés
Cuántos: 140,500 imágenes
Dónde: Bangladés (Daca)
Cuándo: Publicado en 2026
Tipo de estudio: experimento
Quién lo hizo: Universidad del Norte Sur, Daca, Bangladés
El límite que importa: No se probó en una elección real ni en su comunidad

Rostros reales marcados como falsos por error, antes y después de añadir imágenes de Bangladés

Antes de añadir imágenes de Bangladés22%

Después de añadir imágenes de Bangladés10%

Porcentaje de rostros reales del sur de Asia marcados como falsos por error, medido sobre 100 rostros reales reservados para la prueba; la muestra es pequeña y no representa a toda la región.

El mismo modelo, con y sin imágenes reales de Bangladés en el entrenamiento

Modelo sin imágenes del sur de Asiafrente aModelo con 400 imágenes reales de Bangladés añadidas

Los falsos positivos bajaron de 22.0% a 10.0% y el Brier score mejoró de 0.180 a 0.110

Un rostro falso está construido para parecer real, no para resistir una verificación.
Lectura de weeklyAI

Así podría verse · ilustración generada por weeklyAI.watch, no es una fotografía

1. Un equipo de la Universidad del Norte Sur, en Daca, Bangladés, entrenó un sistema de cómputo para distinguir caras reales de caras falsas hechas con inteligencia artificial. Combinó 140,000 imágenes públicas con 500 fotos reales de personas bangladesíes. En sus pruebas acertó el 99% de las veces, y al añadir 400 imágenes del sur de Asia los falsos positivos bajaron de 22% a 10% sobre 100 caras reservadas. Todo esto ocurrió en un laboratorio, no en una elección ni en redes sociales en vivo.

2. Las 500 fotos bangladesíes no son públicas por privacidad, así que usted no puede repetir la prueba con caras de su país. El documento no nombra ninguna página de autoridad electoral, registro o verificador que usted pueda consultar esta noche.

3. El artículo no menciona ninguna regla electoral ni resolución de autoridad. No hay norma citada ni auditoría de una elección.

Qué significa para usted

Para su voto, este estudio no cambia nada todavía: fue una prueba de laboratorio con imágenes de Bangladés, no una elección ni su comunidad, y las fotos usadas no son públicas. Cuando lea sobre herramientas así, pregunte qué autoridad electoral las autorizó, qué auditaron y qué encontraron; hasta entonces, no concluya que su voto o el conteo ya están protegidos por este hallazgo.

Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099

Quién pagó: Los autores no recibieron financiación específica para este trabajo, y el artículo no indica que ningún financiador, proveedor de equipos o de software haya tenido influencia en el diseño, el análisis o la redacción.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1086 palabras · unos 5 minLeerla →Cerrar

La regla y la evidencia: herramienta de laboratorio detecta caras falsas con 99% de acierto, pero no se probó en una elección real ni en su comunidad

Un estudio midió cuántos rostros falsos detecta una herramienta entrenada con imágenes de Bangladesh. Lo que no midió es su rostro, su país ni su elección.

### EL ACTO

Esta semana no hay un acto que citar. Lo que hay es un estudio publicado en una revista científica, y su texto dice qué hizo y con qué límites: "Our study introduces an approach, called ARC-Net"¹. Un estudio no es una resolución, no está en vigor para nadie, no lo aplica ninguna autoridad y no fija plazo alguno. Nadie queda obligado por él, y ninguna persona puede invocarlo para reclamar nada.

Lo único que el documento promete es hacia adelante: "This study will help advance DF detection by integrating ARC-Net's attention-residual mechanisms and XAI, offering insights for developing models in security and media forensics"². Eso es una expectativa de los autores, escrita por ellos, no una norma.

### EL CASO QUE RESPONDE

Esta semana no se encontró ningún caso real que contar. No hay un video con nombre, un lugar y una fecha, ni una segunda fuente que lo sostenga, y por eso esta parte se escribe en una sola frase.

### LO QUE SE MIDIÓ

La muestra fue de 140,500 imágenes: un conjunto público de rostros reales y falsos al que los autores añadieron 500 rostros reales de personas de Bangladesh. El estudio corrió durante 30 épocas de entrenamiento.

La herramienta acertó en el 99% de las imágenes de prueba, con precisión de 1,0, recall de 0,97 y F1 de 0,98³. En otros dos conjuntos públicos, el mismo modelo llegó al 97.60% y al 99.33%⁴.

La parte que más importa para un votante: al añadir menos del uno por ciento de rostros reales de Bangladesh, la tasa de falsos positivos —rostros verdaderos marcados como falsos— bajó "from 22.0% to 10.0%" y el Brier score mejoró "from 0.180 to 0.110"⁵. Para ponerlo en algo que se ve: de cada cien rostros reales de esa región, veintidós se marcaban como falsos antes, y diez después.

La prueba estadística se hizo sobre 100 rostros reales del sur de Asia reservados para eso, y dio un valor de 0,028⁶. Cien rostros es una muestra pequeña para hablar de un grupo entero.

El estudio no se midió en ninguna elección. No hubo votantes decidiendo, ni mesas, ni actas, ni conteos. Fue una evaluación controlada con imágenes ya reunidas.

Y hay un límite que conviene decir sin rodeos: "The 500 images that used to create hybrid dataset in this study is not publicly available due to ethical and privacy concerns. Other datasets are publicly available"⁷. Los rostros de Bangladesh que hicieron posible la mejora no están a la vista de nadie.

### LA RESPUESTA DE LA OTRA REGIÓN A LO MISMO

Esta semana no encontramos ningún tribunal, ley ni autoridad de otro país que se haya pronunciado sobre este método.

### LA LECTURA DE LA REDACCIÓN

Así lo leemos nosotros. Un rostro falso está construido para parecer real, no para resistir una verificación. Por eso la nitidez de una imagen no prueba nada: lo bien hecho y lo verdadero se parecen demasiado. De ahí esperamos algo concreto en las casas como la suya: que una cara falsa bien armada pase por cierta ante usted, y que una cara verdadera —la suya, la de su vecina, la de una candidata— sea señalada como falsa, sobre todo si su rostro, su tono de piel o su región no estaban entre los ejemplos con que se entrenó la herramienta. Sabríamos que nos equivocamos si la herramienta acertara igual en rostros de todas las regiones y tonos, y si sus errores no se concentraran en un grupo; o si el propio estudio mostrara tasas de error parejas por subgrupo en lugar de una mejora solo después de añadir imágenes locales. Usted puede hacer algo con esto desde hoy: ante un video o una foto que parezca decisiva, no la trate como prueba por su calidad. Busque quién la publicó primero, cuándo y de dónde salió. Y pregunte a su autoridad electoral si las herramientas de verificación que usa fueron probadas con rostros y voces de su propia población.

También lo leemos así. Los sistemas que ordenan y deciden lo que vemos se presentan como neutrales porque se apoyan en números, pero funcionan de manera opaca y no siempre aciertan; quien sale perjudicado no alcanza a entender ni a contestar la decisión. De ahí esperamos que una herramienta de este tipo llegue a los medios, a las campañas o a una mesa como si fuera un veredicto automático, sin que el votante común sepa con qué imágenes se entrenó, quién la opera ni cómo reclamar si lo señala mal. Sabríamos que nos equivocamos si la herramienta publicara sus datos de entrenamiento, sus tasas de error y un canal abierto de reclamo, y si sus resultados pasaran por personas antes de usarse contra alguien. Lo que usted puede hacer es sencillo: cuando alguien le muestre un detector de falsos como árbitro final, pida saber quién lo hizo, con qué imágenes se entrenó y cuánto falla. Un porcentaje alto de acierto no le dice nada sobre su propio caso.

### LO QUE NO SE SABE

El estudio probó la herramienta en otros conjuntos públicos distintos de los de entrenamiento, y en esas pruebas mantuvo un rendimiento alto y equilibrado.

No se sabe si mejoraría la detección en un uso real, porque fue una evaluación controlada y no una prueba en condiciones de campaña o de mesa.

No se sabe si reduciría el acoso o la desinformación en la práctica, porque esos efectos no se midieron.

No se sabe si supera a todos los demás métodos, porque solo se comparó con cuatro modelos y con variantes del propio diseño.

No se sabe si el modelo es justo entre grupos, porque solo se evaluó el subgrupo del sur de Asia.

No se sabe si añadir un puñado de rostros de su país arreglaría los falsos positivos en su comunidad, porque el estudio no lo probó en ninguna otra población.

No se sabe si otra persona replicó estos resultados, porque no se reporta ninguna réplica independiente ni auditoría externa.

Ante una foto o un video que parezca decidir su voto, pregunte primero quién lo publicó y cuándo; y pida a su autoridad electoral que le diga con qué rostros se probaron sus herramientas de verificación.

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Our study introduces an approach, called ARC-Net, which uses a combination of attention and residual convolutional layers along with the EfficientNet B0 base, the attention mechanism of which allows the model to pay more attention to details that could resemble the ones seen in non-perfectly executed DF"
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This study will help advance DF detection by integrating ARC-Net’s attention-residual mechanisms and XAI, offering insights for developing models in security and media forensics."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "ARC-Net performed much better than the other traditional and state of art methods with 99% accuracy, 1.0 precision, 0.97 recall and 0.98 F1 score, reaching the highest level of reliability in spotting DF images."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "To assess the external reliability and generalizability of ARC-Net, the model was evaluated on the Deepfake Dataset and the Deepfake Database datasets, achieving consistently high and balanced performance across different scales."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This inclusion reduced the false positive rate from 22.0% to 10.0% and improved Brier score from 0.180 to 0.110."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "statistical analysis using the McNemar’s test yielded a p-value of 0.028, indicating statistically significant results on 100 held-out real South Asian images."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The 500 images that used to create hybrid dataset in this study is not publicly available due to ethical and privacy concerns. Other datasets are publicly available."

experiment · PloS one · the paper, 6 Apr 2026 · free

A deepfake detector for faces was tested in a lab, not in an election

Short version · the longer version follows, about 7 min

Ask weeklyAI

Ask me about this study: who was studied, what it found, and what it does not say.

Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.

The study at a glance
Who: real and fake face images, including 500 real faces of Bangladeshi people
How many: 140,500 images
Where: Bangladesh
When: published April 6, 2026
Kind of study: experiment
Who did it: North South University, Dhaka, Bangladesh
The limit that matters: Tested only on still images in a lab, never in an election or on a live feed.

How often the detector called a real local face fake, before and after adding local faces

Before adding local faces22%

After adding local faces10%

These are the shares of real South Asian faces wrongly called fake, measured on 100 held-out real images; the paper says the change was statistically significant but the test set was small.

Detector accuracy on the main mixed set versus two outside sets

Main mixed set of 140,000 public faces plus 500 Bangladeshi facesagainstDeepfake Dataset

99% accuracy on the main set, 97.60% on the outside set

Main mixed set of 140,000 public faces plus 500 Bangladeshi facesagainstDeepfake Database

99% accuracy on the main set, 99.33% on the outside set

The article names no way for a voter to check a suspicious video or message before voting.
weeklyAI's reading

How it could look · illustration generated by weeklyAI.watch, not a photograph

Researchers at North South University in Dhaka, Bangladesh, published on April 6, 2026, a test of a tool called ARC-Net that sorts real face images from manipulated ones. They trained and tested it on 140,500 images, of which 500 were real faces of Bangladeshi people collected with consent. The article reports 99% accuracy on that mix. The 500 Bangladeshi images are not public, and no real election, live feed or voter register was involved. The article states manipulated face images are used in South Asia to harass women and attack female politicians.

The article names no way for a voter to check a suspicious video or message before voting. It points to no register page, ad library or provenance mark for you to open tonight.

The article records no electoral authority rule, ruling or audit on this tool. It states the study was approved by the North South University ethics committee, application 2024/OR-NSU/IRB/1109, and that the 500 images stay confidential.

What this means for you

When you see a clipped face or a too-smooth video before voting, remember this study was a lab test, not a check of any live feed or register. It names no page, mark or audit for you to verify tonight, so weigh what you can find and ask your electoral authority what it allowed and found.

Who paid: The authors received no specific funding for this work, and the article does not say that any funder, lender or software provider had a say in the design, analysis or reporting.

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1351 words · about 7 minRead it →Close

The rule and the evidence: A deepfake detector for faces was tested in a lab, not in an election

One team in Bangladesh trained a detector on 500 local faces. The gain was real and narrow. The rule that would put it in your polling place doesn't exist yet.

### THE ACT

No binding rule, resolution, ruling or bill on deepfake detection appears in the record this week. What exists instead is a research paper, published April 6, 2026, describing a model called ARC-Net, which its authors built to tell real faces from fake ones. Nothing in it is in force anywhere, binds anyone, or reaches a polling station.

The paper describes a tool, not a law: "Our study introduces an approach, called ARC-Net, which uses a combination of attention and residual convolutional layers along with the EfficientNet B0 base"¹. That is the whole of what was adopted here — a method, tested in a lab, offered to the field of media forensics².

What a person can do with it today is nothing direct. The 500 Bangladeshi faces that gave the model its local edge are not available: "The 500 images that used to create hybrid dataset in this study is not publicly available due to ethical and privacy concerns. Other datasets are publicly available"³. There is no authority to petition, no date to act by, and no enforcement behind any of it.

### THE CASE IT ANSWERS

No real instance was found this week — no specific deepfake with a name, a place, a date and a second source.

The record describes the class of harm without naming a single case: manipulated facial images in Bangladesh and South Asia are "primarily utilized to produce content, circulate misleading information, and instill feelings of fear or disgust among the people"⁴. It would seem this is the kind of thing they mean.

The paper goes further, and again without a name: "Because of the way AI is misused to target people with harassment, thousands of women have had to take down their pictures and videos from social media sites"⁵. And: "this deepfake technology is also used to increasingly attack female politicians in South Asia"⁶. No victim is named, no date given, no outlet cited here.

For anything short of a documented case, the record is short. There is no incident in it that a reader could look up, no platform, no month, no second source — only the general statement that the problem is visible in Bangladesh⁷. The fake itself is never reproduced, and this story does not reproduce one either.

### WHAT WAS MEASURED

ARC-Net was tested against four older models and reported 99% accuracy, 1.0 precision, 0.97 recall and a 0.98 F1 score⁸. On a test set of 20,100 images it called about 99 in every 100 correctly; the precision figure means that when it says fake, it is essentially never wrong, and the recall figure means it misses about three fakes in a hundred.

The sample: a hybrid set of 140,000 public faces plus 500 Bangladeshi faces, tested on 20,100 images, with the fairness question checked on 100 held-out real South Asian images⁹¹⁰. It was measured in a lab in Bangladesh, on still images — never in an election, never on a live feed.

Adding fewer than one percent real Bangladeshi images cut the false positive rate from 22.0% to 10.0% and improved the Brier score from 0.180 to 0.110 — meaning the model stopped calling roughly one in nine real local faces fake instead of one in five¹¹⁹. The improvement was statistically significant by McNemar's test, p = 0.028, on those 100 images¹⁰.

Two cross-dataset checks held up less brightly, and the paper says so: 97.60% accuracy on one outside dataset and 99.33% on another¹². An ablation showed that combining residual and attention modules added 5.0% accuracy over the bare baseline, and removing either piece cost ground¹³.

Across 36 studies and 13,197 participants, people performed above chance at 56.1%, with higher accuracy for real faces than fake ones, and wide variation between studies¹⁴¹⁵¹⁶. The same abstract notes that experiments trying to improve detection reached 62.2% but that the practical relevance of that should be treated with caution¹⁷.

Other detection work in the record is weaker still under cross-dataset conditions: one comparison of four video models found the best reaching only 64.68% test accuracy, with "significant degradation in generalization"¹⁸¹⁹. The gap between a lab score and a strange dataset is the whole story of this field so far²⁰.

### THE OTHER REGION'S ANSWER TO THE SAME THING

We found none this week — no court, law or authority outside Bangladesh is named in the record as answering this same method.

### THE LIBRARY'S READING

Here is how we read it. A false story about a candidate has always worked before the ballot, and it has always worked the same way: it arrives faster than the correction, it leaves a residue that no retraction clears, and it asks the voter to do the one thing nobody can do reliably — tell, in a glance, whether the face in front of them is the person they think it is. What is new is not the lie. What is new is that the lie can now wear a face, and that the face can be produced in the time it takes to open an app. What a public can learn to check is the small, boring things: whether a second outlet that does not share your politics has the same image, whether the file existed before the date it claims. What a public cannot check alone is whether the tool that flags the image was ever tested on faces like theirs. That is not a skill a voter acquires. It is a question only an institution can answer.

So we expect this in homes like yours: a suspicious picture will arrive, someone will look at it, and the looking will settle nothing — because the person will see what they already believed about the side shown. We would be wrong if it turned out that people can reliably spot a manipulated face even when they are primed to believe it, and that their mistakes do not cluster around the politicians they already distrust. What you can do with this is small and usable. When an image about a candidate reaches you before a vote, do not ask whether it looks real. Ask who else has it, and whether anyone who does not already agree with you has confirmed it. If a tool is ever used to check faces at a polling place or in a campaign, ask who tested it on your community's faces — and treat its answer as a lead, not a verdict.

### WHAT IS NOT KNOWN

The record is silent on whether any of this reaches a voter. The 99% figure comes from a still-image test set, not a live feed, not an election, and the paper itself frames the work as an opening for future research rather than a deployment².

The fairness gain rests on 100 held-out real South Asian images and a single research group's model — no independent replication, no third-party audit appears¹⁰³. The paper never claims the tool is fair across all groups; only that subgroup was checked.

And the local faces that produced the gain cannot be examined by anyone outside the team, because they are not public³. What is not known is whether the same small addition of faces would fix the same problem for a community in Latin America, the United States or Canada. The record does not say, and this story will not pretend it does.

One more thing we cannot know from here: whether a tool like this, if it ever sat between you and your ballot, would be pointed at the fake or at the real face that happens to look wrong. You will not find that out from a paper. You will find it out by asking, before the date fixed by law, who is checking the images — and on whose faces that check was ever tested.

Where each piece of context comes from, and how much of it we read

Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "Our study introduces an approach, called ARC-Net, which uses a combination of attention and residual convolutional layers along with the EfficientNet B0 base, the attention mechanism of which allows the model to pay more attention to details that could resemble the ones seen in non-perfectly executed DF"
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "This study will help advance DF detection by integrating ARC-Net’s attention-residual mechanisms and XAI, offering insights for developing models in security and media forensics."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "The 500 images that used to create hybrid dataset in this study is not publicly available due to ethical and privacy concerns. Other datasets are publicly available."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "In regions like Bangladesh and other countries in South Asia, manipulated facial images are primarily utilized to produce content, circulate misleading information, and instill feelings of fear or disgust among the people."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "Because of the way AI is misused to target people with harassment, thousands of women have had to take down their pictures and videos from social media sites."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "Also, this deepfake technology is also used to increasingly attack female politicians in South Asia [ 6 ] and [ 7 ]."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "Deepfake (DF) content poses a major challenge to digital media authentication, that can mimic facial movements, creating realistic replicas that risk spreading misinformation and enabling harassment, as can be seen in Bangladesh."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "ARC-Net performed much better than the other traditional and state of art methods with 99% accuracy, 1.0 precision, 0.97 recall and 0.98 F1 score, reaching the highest level of reliability in spotting DF images."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "This inclusion reduced the false positive rate from 22.0% to 10.0% and improved Brier score from 0.180 to 0.110."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "statistical analysis using the McNemar’s test yielded a p-value of 0.028, indicating statistically significant results on 100 held-out real South Asian images."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "the first one evaluated South Asian images, showing that incorporating fewer than one percent real Bangladeshi images reduced the false positive rate by more than half and improved probability calibration"
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "To assess the external reliability and generalizability of ARC-Net, the model was evaluated on the Deepfake Dataset and the Deepfake Database datasets, achieving consistently high and balanced performance across different scales."
Reza, M. S., Elias, F., Mahmud, M. I. et al. (2026). Attention and residual mechanism-based CNN architecture (ARC-Net) with enhanced fairness generalization for deepfake facial image detection. PLOS One. https://doi.org/10.1371/journal.pone.0340099 - the article this story is about — the whole article — the passage: "An ablation study further showed the impact of different components within a model by systematically removing or modifying them and statistical significance between competing classifiers was assessed using McNemar’s Statistical test."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "Including 36 studies (k = 51 experiments), totaling 13197 participants, we tested if humans perform above chance-level in face detection tasks and if accuracy is higher for real (vs. deepfake) stimuli, alongside moderator analyses including a series of methodological variables (i.e., publication year, study quality, deepfake type and response modality)."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "Overall, our results provide evidence that humans perform above chance-level (56.1%), and confirm significantly higher accuracy for real vs. deepfake face stimuli."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "However, significant heterogeneity observed suggests high variability in the results obtained by individual studies and both effects are shown to be moderated by the assessed variables."
Stockner M, Convertino G, Cambedda S, Mazzoni G. (2026). Are humans able to discriminate between real and deepfake faces? A systematic review and meta-analysis. Computers in Human Behavior: Artificial Humans. 10.1016/j.chbah.2026.100332 — only the abstract - the full text could not be fetched — the passage: "Finally, sub-group analyses on experiments aiming to improve deepfake detection indeed reached considerably higher accuracy (62.2%). However, practical relevance of the encountered accuracy should be considered with caution."
Agrawal P, Pathak D, Madaan V, Verma PK, Choo WO. (2026). Spatiotemporal deep learning for real-time video-based deepfake detection using 3DCNN, 3DResNet, TCN, and VAE. Scientific Reports. 10.1038/s41598-026-49090-1 — only the abstract - the full text could not be fetched — the passage: "The experimental findings indicate that 3DCNN can reach the highest test accuracy of 64.68% which is higher than the results of 3DResNet, TCN, and VAE under the cross-dataset conditions."
Agrawal P, Pathak D, Madaan V, Verma PK, Choo WO. (2026). Spatiotemporal deep learning for real-time video-based deepfake detection using 3DCNN, 3DResNet, TCN, and VAE. Scientific Reports. 10.1038/s41598-026-49090-1 — only the abstract - the full text could not be fetched — the passage: "It analyses indicate that there is significant degradation in generalization and different failure behavior with models when presented with heterogeneous distributions of data."
Shivaprakash SJ, H S, Chauhan A, Md AQ. (2026). Attention-augmented hybrid framework with evolutionary optimization for robust deepfake detection. Scientific Reports. 10.1038/s41598-026-51284-6 — only the abstract - the full text could not be fetched — the passage: "While various deepfake detection models have been proposed, many struggle to maintain consistent performance across datasets and unseen video formats, revealing a clear research gap in generalizability and temporal feature modeling."

Who paid: The authors received no specific funding for this work, and the article does not say that any funder, lender or software provider had a say in the design, analysis or reporting.