Introducing
Neural Defend's foundational model for deepfake detection. ARCAS decomposes a single frame into spectral, noise, compression and geometric layers, inspects each one, and fuses the evidence into one calibrated score.
[MODULE 01//Concept labPreview]
C01 // Anatomy of a deepfake
What does a single synthetic frame look like when every forensic stratum is separated?
Pixel grid, noise floor, frequency starburst, geometry mesh and heatmap. Each stratum holds evidence the others miss.
ARCAS 1BARCAS separates these strata in a single pass and weighs them together.
C02 // Generator fingerprints
Can one model learn the signatures of generator families it was never shown?
GAN lattices, diffusion grain, face-swap seams and tiled decoders all leave distinct traces. ARCAS reads them into one shared representation.
ARCAS 1BARCAS learns generator traces as one representation, so a new tool still leaves evidence it can read.
C03 // Attack surface
Where does synthetic media enter a real workflow, and where should detection sit?
Camera injection, replayed uploads and cloned voices arrive through different channels. Intercepting them at one scan gate keeps policy consistent.
ARCAS 1BARCAS sits at that gate and reads every channel with one foundation model.
C04 // Voice spectrogram
Where in the spectrum does a cloned voice stop sounding human?
Formant ridges, breath noise and harmonic texture form a landscape. Synthetic segments flatten or repeat in ways the ear forgives.
ARCAS 1BARCAS reads the voice spectrum for the textures cloning leaves behind.
C05 // Forensic autopsy
What does a full examination of one face reveal, layer by layer?
Specular highlights, skin micro-texture and blend boundaries are lifted off and examined separately, the way an analyst works a case in NeuroDetect.
ARCAS 1BARCAS runs this examination automatically, on every frame it receives.
C06 // Latent space
In the space a model learns, how far apart are real and synthetic faces?
Real photographs gather on one manifold, generated faces on another. Detection is drawing the boundary between them and keeping it sharp as generators move.
ARCAS 1BARCAS is forensically trained to keep that boundary sharp as generators move.
C07 // The mirror test
When a synthetic twin looks right at a glance, which physical cues give it away?
Specular reflections, light direction and micro-timing must agree with the scene. A synthetic twin often matches the face but not the physics.
ARCAS 1BARCAS checks the physics of the scene, not just the face.
C08 // Temporal forensics
How do you find the single frame where a video stops being consistent?
Identity, lighting and motion should drift smoothly from frame to frame. A manipulated frame breaks the sequence with flicker and temporal drift.
ARCAS 1BARCAS tracks consistency across frames to find the one that breaks.
[MODULE 02//Forensic researchActive research]
R01 // Representation learning
Can a single backbone learn a forensic representation that transfers across image, video and audio, instead of one brittle classifier per media type?
We study traces that survive a change of carrier: specular consistency in the eye, skin micro-texture, and noise statistics that hold across frames.
ARCAS 1BARCAS is that backbone: one foundation model behind every check.
R02 // Out-of-distribution generalisation
Does the decision boundary hold for samplers, VAE decoders and fine-tunes that were absent from training?
Each generator family leaves a different fingerprint: GAN upsampling grids, diffusion denoising residue, decoder smoothing. The research target is the trace, not the tool.
ARCAS 1BARCAS is trained on the trace, not the tool, so new generators still leave something it can read.
R03 // Adversarial robustness
How stable is p(synthetic) under recompression, resizing, colour filters and gradient-based evasion that targets the detector itself?
Recompression wipes high-frequency detail, filters shift colour statistics, and targeted noise attacks the model directly. We measure which forensic signals survive each transform.
ARCAS 1BARCAS is trained to hold its verdict when an attacker reshapes the pixels.
R04 // Audio-visual consistency
When lip kinematics, voice timbre and facial dynamics disagree, how should the evidence be fused and how should the conflict be reported?
Cloned voices carry artifacts in formants and breath; lip-sync models drift between phoneme and viseme timing. Disagreement between channels is evidence in itself.
ARCAS 1BARCAS weighs voice and face together and reports exactly where they disagree.
R05 // Capture-loop inference
Can forensic-grade inference run inside a live capture flow, before the user notices the wait?
Injected camera streams, virtual devices and replayed frames reach the pipeline before any face is analysed. Our product target is under 200 ms p95 on eligible image requests.
ARCAS 1BARCAS runs inside the capture flow, before the decision is made.
[MODULE 03//Threat taxonomyMapped]
One identity pasted onto another person’s video.
ARCAS readsBlending boundaries, identity-texture mismatch
A real face driven by someone else’s expressions.
ARCAS readsLandmark dynamics, temporal consistency
A face or scene generated from nothing.
ARCAS readsGenerator fingerprints, spectral peaks
One region regenerated inside a real photo.
ARCAS readsLocal noise, compression-history mismatch
Synthetic speech in a real person’s voice.
ARCAS readsFormant and breath artifacts, lip-sync drift
A fake stream fed in place of the live camera.
ARCAS readsCapture-path signals, replayed-frame traces
[MODULE 04//Open problemsActive research]
Finding a small inpainted region inside an otherwise authentic photograph, where most of the evidence says “real”.
New generators and community fine-tunes ship every week. The trace has to generalise faster than the tools change.
Attackers who know a detector exists optimise against it. Robustness to targeted evasion is an open problem for the whole field.
Phone-quality audio, low light and heavy compression degrade every channel at once. Fusing weak signals honestly is unsolved.
A score should mean the same thing on every device, codec and platform. Keeping it calibrated outside the lab is ongoing work.
[MODULE 05//Research notesLab notes]
Generators that upsample learn to hide it in pixel space, not in the frequency domain. Periodic structure from transposed convolutions shows up in the Fourier magnitude as a grid of bright points the eye never sees.
Every platform re-encodes uploads. The high-frequency traces a simple detector relies on are the first thing JPEG quantisation removes, so robust detection has to rest on signals that survive the round trip.
Lip-sync models align mouth shapes to audio frame by frame. Small timing offsets and coarticulation errors accumulate, and the mismatch between what is heard and what is seen becomes measurable evidence.
[MODULE 06//PublicationsPreprint]
Evaluates ARCAS 1B across seven public benchmark families without benchmark-specific parameter updates. It reports each benchmark's native metrics alongside pooled results, record-level coverage accounting and subgroup diagnostics, and restricts comparisons with prior work to aligned evidence. The authors state that the findings cover the evaluated records only.
[MODULE 07//Model cardPublished]
[MODULE 08//CollaborateOpen call]
Open call // Research partners
We work with universities, red teams and platform trust and safety teams on datasets, adversarial testing and shared research questions.
Start a research conversation