LC Wanderlust
Fabrizio Musacchio#Deepfakes are everywhere, but #DigitalForensics investigators are fighting back:
Aude Caussarieu đ»Ah bah voilĂ une super source d'exos d'optique gĂ©omĂ©trique đ
Mathieu@AudeCaussarieu
J'aurais jamais pensĂ© que les points de fuites pouvaient ĂȘtre une façon de vĂ©rifier la vĂ©racitĂ© d'une image. đ€Ż
J'aurais jamais pensĂ© que les points de fuites pouvaient ĂȘtre une façon de vĂ©rifier la vĂ©racitĂ© d'une image. đ€Ż
Mourioche â@mathieu_ @AudeCaussarieu
Ăa, ça marchera jusqu'Ă ce que les IA gĂ©nĂ©ratives d'images se mettent Ă placer les Ă©lĂ©ments d'une image dans un espace en 3D.
Ăa, ça marchera jusqu'Ă ce que les IA gĂ©nĂ©ratives d'images se mettent Ă placer les Ă©lĂ©ments d'une image dans un espace en 3D.
@mathieu_ @AudeCaussarieu Je mets ma main au feu que les IA gĂ©nĂ©ratives savent dĂ©jĂ le faire, si on leur demande dans le prompt đ
On va surement en arriver au mĂȘme rĂ©sultat que pour la presse ou la tĂ©lĂ©vision : La confiance dans le mĂ©dia sera liĂ© Ă la rĂ©putation du mĂ©dia et Ă rien d'autre...
GĂ©rard Fadrelle@Mourioche bah j'avais vu une vidĂ©o sur la musique gĂ©nĂ©rĂ© par IA oĂč on disait que le gros problĂšme c'Ă©tait que l'IA n'arrive pas Ă reconstituer les pistes/instrument et manipule le bazar comme un tout, et forcĂ©ment qqn d'exercĂ© arrivait Ă repĂ©rer des Ă©trangetĂ© dans les sons et la stĂ©rĂ©o.
C'est le mĂȘme problĂšme pour la gĂ©omĂ©trie, vu qu'on crĂ©e l'image comme un tout, sans "calque" (qui permet de limiter ce genre de couilles en structurant les plans)
---> []@mathieu_ @AudeCaussarieu
Je suis quand mĂȘme un chouilla rĂ©servĂ© sur la premiĂšre photo. Ok, elle semble gĂ©nĂ©rĂ©e par IA, mais les lignes de fuite sur lesquelles l'analyse se base sont un peu courtes. Difficile d'extrapoler sur un petit segment qui n'est pas nĂ©cessairement reprĂ©sentatif ni exact.
La méthode est intéressante mais pas toujours applicable
Je suis quand mĂȘme un chouilla rĂ©servĂ© sur la premiĂšre photo. Ok, elle semble gĂ©nĂ©rĂ©e par IA, mais les lignes de fuite sur lesquelles l'analyse se base sont un peu courtes. Difficile d'extrapoler sur un petit segment qui n'est pas nĂ©cessairement reprĂ©sentatif ni exact.
La méthode est intéressante mais pas toujours applicable
frankiezafe@YannC @mathieu_ @AudeCaussarieu je crois que la chaĂźne qui rentre dans le bras gauche du soldat sur la gauche est un meilleur indice de fake que la perspective inexacte :)
DrYak@mathieu_ @AudeCaussarieu fondamentalement, les AI generatives fonctionnent a l'echelle des pixels d'une image (ou video), et n'ont pas vraiment un model abstrait d'une scene (pas de plan 3D en tete)
C'est pour ca aussi sur les model plus vieux les doigts surnumeraires, les 3e mains mal placees, etc. Meme-si ces erreures la sont plus facilement corrigibles.
Dr Grototo@AudeCaussarieu Je me demande quand mĂȘme Ă quel point câest difficile de distinguer de telles anomalies de distorsions optiques causĂ©es par la lentille, qui va courber les lignes parallĂšles.
ChÚvre Mousse@grototo @AudeCaussarieu On verrait un autre genre de déformation, pas des trucs au pif comme ça, je dirais ? De toute façon dans le genre de photo montré est rarement prise en grand angle...
Nartagnan@legendarybassoon @grototo @AudeCaussarieu
Oui, je suppose ça ferait une forme pas exacte, mais pas random.
L'images des cubes par exemple, il y aurait de la symĂ©trie, pas juste un des points qui va n'importe oĂč.
Et j'imagine la gueule des prompts pour dire de se concentrer sur les points de fuite, difficile Ă dĂ©crire en langage "naturel". Avec un peu de chances ça va pas ĂȘtre facile Ă corriger.
youen@nartagnan @legendarybassoon @grototo @AudeCaussarieu je suppose qu'on est dĂ©jĂ entrĂ© dans une nouvelle guerre "gĂ©nĂ©ration de fakes crĂ©dibles" vs "dĂ©tection de fakes", comme on a dĂ©jĂ les crawlers/rate-limiter, pub/ad-blocker, tracker/anti-tracker, etc. đ
Donc oui, mĂȘme si ce n'est pas facile (?), je ne serais pas surpris que les entreprises derriĂšre ces gĂ©nĂ©rateurs d'images analysent les techniques de dĂ©tection et fassent leur possible pour les rendre inopĂ©rantes.
@youen @legendarybassoon @grototo @AudeCaussarieu
Mais ça va coûter.
MĂȘme avec de la dĂ©tection, s'ils veulent faire comme le claude code, aka "gĂ©nĂšre, on check, et tant que ça passe pas le check recommence, mĂȘme si c'est 1000 fois", ça va leur coĂ»ter... 1000 fois plus cher.
Et ils n'auront pas ce luxe encore trĂšs longtemps.
tk đ@nartagnan l'uncanny valley dans 10 000 ans, si on est encore dans les parages : pourquoi la reconnaissance instinctive des lignes de fuite est devenue un enjeu de sĂ©lection naturelle ?
@nartagnan en fait, je vois mĂȘme pas comment intĂ©grer ça au process d'entrainement, sans que cela devienne une machine Ă gaz, ce qui est dĂ©jĂ le cas however, genre encoder un raytracer
Youen PĂ©ron@tk @nartagnan @legendarybassoon @grototo @AudeCaussarieu
GĂ©nĂ©rer plein d'images par IA, demander a des petites sous payĂ©es de dessiner les lignes fuites. On fait deux jeux de donnĂ©es : les images avec un seul point d'intersection et les autres. On rajoute des vrais images dans la premiĂšre catĂ©gorie. On lance l'entraĂźnement dâun modĂšle ou un fine tunning dâun modĂšle existant.
@youen
@tk @legendarybassoon @grototo @AudeCaussarieu
Oui, c'est faisable.
Mais se concentrer sur X c'est délaisser Y.
Au début, quand il fallait compter les doigts des mains, les modeles qui étaient bons sur les mains étaient mauvais sur le reste.
L'amélioration n'est venue qu'en multipllant le nb de paramÚtre des modÚles. Et donc le coût de génération d'une seule image.
C'est exponentiel.
Et j'ose croire qu'il n'y a plu moyen de multiplier encore par 2 leurs coûts, sans revenus.
@grototo
C'est sĂ»rement plus chiant Ă vĂ©rifier mais il faudrait tracer les cercles et ceux-ci devrait avoir la mĂȘme intersection.
@AudeCaussarieu
C'est sĂ»rement plus chiant Ă vĂ©rifier mais il faudrait tracer les cercles et ceux-ci devrait avoir la mĂȘme intersection.
@AudeCaussarieu
The Marvelous Mx Tellington@AudeCaussarieu oh wow faut que je bosse avec des profs de maths pour faire des séances maths / éducation aux médias !
pulsation
Mans R@FabMusacchio How does this method handle lens distortion?
Klara@mansr @FabMusacchio the middle lines should still meet, the outer ones will cross a little bit in an orderly manner. Not the second to the left and the third to the right.
ChookMother đŠđșđŠ@FabMusacchio Interesting. Thanks for posting!
Microblog Castellano@FabMusacchio So if I want to commit a murder, I have years to prepare it and I know the place will be surveilled with cameras, I should pave it with slightly non-parallell tiles, to get a plausible deniability.
L'Ă©grĂ©gore AndrĂ© êêŹ@microblogc @FabMusacchio Forced perspective is your friend: "The Potemkin Stairs in Odesa extend for 142 metres, but give the illusion of greater depth since the stairs are wider at the bottom than at the top."
Light@microblogc @FabMusacchio I was thinking, what if it was just paved a bit wonky.
F4GRX SĂ©bastien@FabMusacchio beautiful analysis even though I'm still suspicious of lens distortions.
Leah Neukirchen@f4grx @FabMusacchio sun rays are parallel, yet they meet at a point...?
Edvin Malinovskis@nCrazed @f4grx @FabMusacchio But this is irrelevant for the technique described? It would work with a non-parallel spot lamp just as well?
@nCrazed @leah @f4grx @FabMusacchio Lines from a shadow through the object casting it meet at the light source. Sunlight being virtually parallel is irrelevant. If the light source isn't point-like, you get blurry shadows.
@nCrazed @leah @f4grx @FabMusacchio If the light source is behind the image plane, these lines will of course converge at an opposite point. The important part is that they all converge. If they intersect every which way, something funny is going on.
@mansr @nCrazed @leah @FabMusacchio yep I agree with that remark, modulo lens distortions. I would love to see the actual convergence on a known good image, for comparison sake.
@f4grx @mansr @nCrazed @FabMusacchio I tried it on two random hallway images from Kagi Search and it worked great. I think most lens distortions are nonlinear, so the problem is rather that the lines are not straight than that they won't meet.
@leah @mansr @nCrazed @FabMusacchio How do you model the curvatures of these non straight lines so you can be sure that they all meet at the same point? đ€
@f4grx @mansr @nCrazed @FabMusacchio I didn't need to, since the lines were straight I assumed lens correction was already done or not necessary.
Marcus MĂŒller@leah agreed, I just wished the technique they described came with that as a warning, because the (obviously generated, "read" the uniform patches) hallway picture shown would be a *prime* candidate for taking with a fisheye lens or a similarly distorting lens; and the piece of flooring used to extrapolate the straight lines is already honestly too short in the example to be sure. I cannot, over the length of maybe 50px, draw a 1000px line with < 1° error.
@leah (and of course, a photojournalist in a government building is more likely to have a "low distortion as possible" lens equipped than a fishlens, but if you're preparing for e.g. reporting from a small, crowded room and want to get as many interacting politicians into the picture as possible, you'd not go in there with a superzoom lens alone. And if a picture is claimed to be from a publicity shot, the photographer certainly will pick from a wide range)
Stanley Coupdebrique@nCrazed @leah @f4grx @FabMusacchio Sunrays are quasi-parallel in the 3rd dimension, however they all diverge from the sun when projected on a surface such as a camera sensor or our retina. What difference does it make if the trick works with lamps too ? The sun is just a big, faraway lamp, and the rules of convergence in 2d spaces apply just the same for all 1d-ish lightsources.
The trick works if you don't mind for lens distortion, which are really minor on most smartphone cameras nowadays (do not mess lens distortion with surface distortion caused by planar projection on sensors, the latter doesn't affect convergence even though it can exxagerate surfaces on the sides of the pic).
@songxisto @nCrazed @f4grx @FabMusacchio Assume a solid cube on a plane, and only parallel, ambient light hits it from an angle. if you connect the cube corners with the three visible shadow-corners, the lines are parallel as well and won't intersect. am i wrong?
@leah @nCrazed @f4grx @FabMusacchio You are right if the image is itself a parallel projection, but the images from cameras are perspectives, where parallel rays/lines in 3d always converge somewhere (except rays/lines parallel to the sensor).
@songxisto @leah @nCrazed @FabMusacchio Yep. you can see the difference in freecad by switching between parallel projection and perspective.
LeaIt's not the sun's rays that meet at a point, it's the lines from the objects' shadows to the corresponding points on the objects that should meet at a point.
The statement about the sun's rays being effectively parallel just means that the direction of the light source can be considered the same for all objects.
Seachaint@leadore @leah @f4grx @FabMusacchio I'm still bothered though: the shadows of those cubes are all splayed-out relative to one another, but in actual sunlight they should be parallel to one another. But the "proof that it's fake" seems to ignore that?
Glenn Hoppe@seachaint @leadore @leah @f4grx @FabMusacchio because the sun is at a distance, parallel lines meet at a vanishing point, like train tracks.
@ghoppe @leadore @leah @f4grx @FabMusacchio While that is technically true, the splay of shadows on Earth is negligible and nearly imperceptible in real life. In the example image it's as if the sun is a lamp a few metres away.
@seachaint @leadore @leah @f4grx @FabMusacchio in most situations, shadow perspective is apparent when the camera is pointing towards the sun. Itâs different if camera is overhead.
@seachaint
Yeah, I noticed that. All the example images have other things about them that are wrong and give them away as fake.
I think the OP wanted to just focus specifically on how to check for lines that should meet at a point, as one kind of objective test we can use, and not get into all the other stuff.
Xauri'EL Zwaan@FabMusacchio it is really telling that one of the most powerful tools to detect fake images is an artistic tool that any first year art school student knows about
Martin Vermeer 86 47@XauriEL @FabMusacchio Yep. GenAI doesn't _understand_ anything - like 3D geometry - it just knows a lot of patterns. The same holds for LLMs. The way to test is to look for understanding, or the lack of it.
Like this
https://bsky.brid.gy/r/https://bsky.app/profile/did:plc:jdgad3rwwpboirmdcqx4qqvm/post/3mjs54ulm622n
Ilusënn@FabMusacchio Hello! Can I ask where the url shortener sends?
Jonathan T@ilusenn @FabMusacchio It's legit - the website of the scientific journal, Science (they picked scim.ag because it is short for Science magazine).
Here's the full URL: https://www.science.org/content/article/deepfakes-are-everywhere-godfather-digital-forensics-fighting-back
Petra van Cronenburg@JonnyT Thanks! For I never click on URL shorteners ... like many people here. @ilusenn @FabMusacchio
@NatureMC @ilusenn @FabMusacchio Definitely not a fan either but I am aware of where this one leads (or, at least, should).