Skip to content
AI360XpertBeta

Synthetic Media Detection

A detector gives every clip a fakeness score and calls it fake past a threshold. You trade false alarms against misses by moving that line.

A detector scores each clip and calls it fake at or above 0.5, giving two hits, two correct clears, one false alarm and one miss across the six worked scores
A detector scores each clip and calls it fake at or above 0.5, giving two hits, two correct clears, one false alarm and one miss across the six worked scores

Why Does This Exist?

Some clips arrive with a verifiable mark of how they were made. Most hostile ones do not. A maker who wants to deceive will never stamp their own fake, so marks only help when the generator cooperates. The full treatment of marks is in watermarking and provenance.

That leaves you with the hard case: a plain clip, no history, and a question you have to answer anyway. Your eyes are no help here. Modern generators have perfected the big picture, faces, rooms, lighting at a glance, while the remaining flaws hide in micro cues you cannot see. Synthetic media detection is the after-the-fact answer: score the clip for generator traces and decide by threshold.

This page covers the detector itself: the score, the line, the two cue families, and one worked example with real numbers. Judging overall media quality lives in generative media evaluation, and how the fakes get made lives in video generation. If scores and thresholds feel fuzzy, the short version is that a threshold turns a ranking into a decision, and precision counts how often your alarms are right while recall counts how many fakes you catch.

Think of It Like This

A bank teller feeling a banknote

A teller barely glances at the printed portrait. Forgers perfect portraits first, so the portrait proves nothing. Instead she feels the paper, tilts the ink, finds the security strip. She ignores the big picture the forger mastered and checks the micro cues the forger finds hardest.

Her other tool is a line in her head. Faint doubt passes, real suspicion calls the manager. Lower that line and she catches more fakes while delaying more honest customers. Raise it and the queue moves but a forgery slips through. That is your threshold.

Where it stops: currency is a fixed set with stable tests. Your opponent reads your research and retrains. Every published cue has a shelf life, so your line needs rechecking in a way hers never does.

How It Actually Works

A detector maps each clip to one number and compares it against a line. Call the clip xx. Call the fakeness score s(x)s(x), a number between 0 and 1 where larger means more synthetic looking. Call the threshold τ\tau, the line you choose. Call the verdict y^\hat{y}, which is 1 for fake and 0 for real. Call the truth yy, what the clip actually is. The whole rule is one line:

y^=1 when s(x)≥τ, else 0\hat{y} = 1 \text{ when } s(x) \ge \tau, \text{ else } 0

What do the score and the line mean?

Every clip lands in one of four outcomes. Say them in plain words first, then keep the short names because the metrics need them:

  • Hit (true positive): truly fake, scored at or above the line. You caught one.
  • False alarm (false positive): truly real, scored at or above the line. You cried wolf.
  • Miss (false negative): truly fake, scored below the line. One slipped through.
  • Clear (true negative): truly real, scored below the line. Correctly left alone.

Move τ\tau left and hits rise with false alarms. Move it right and false alarms fall with hits turning into misses. That exchange is the entire job. There is no position that removes both errors, only the split your setting can live with.

Where does the score come from?

The number s(x)s(x) blends two cue families, and both matter because each fails on different fakes:

  • Body and physics consistency. Generators arrange pixels without living in a body. So detectors check blink patterns, subtle skin pulse signals, whether reflections and shadows agree on one light source, and whether teeth, fingers and edges hold together. One odd cue rarely convicts. Several together push s(x)s(x) up.
  • Statistical generator traces. Generators build images through learned upsampling, which leaves repeating structure in the frequency spectrum that cameras do not produce. Face swaps add blending seams where pasted pixels meet real ones. Diffusion models leave their own version of these traces, and compression partly smears them, which is why laundered clips score lower.

Neither family reads the scene for meaning. Both ask whether the pixels carry the fingerprints of a camera or of a generator.

Worked example: six clips, one threshold

Take six clips with scores and truths you know:

  • Clip A: s=0.92s = 0.92, truly fake. Clip B: s=0.81s = 0.81, truly fake. Clip C: s=0.55s = 0.55, truly real. Clip D: s=0.48s = 0.48, truly fake. Clip E: s=0.30s = 0.30, truly real. Clip F: s=0.12s = 0.12, truly real.

Set τ=0.5\tau = 0.5. Clips A, B and C sit at or above the line, so y^=1\hat{y} = 1 for those three. Clips D, E and F sit below it, so y^=0\hat{y} = 0 there. Counting against truth: hits are A and B, so TP=2TP = 2. The false alarm is C, so FP=1FP = 1. The miss is D, so FN=1FN = 1. The clears are E and F, so TN=2TN = 2.

Now the rates everyone quotes:

precision=TPTP+FP=23≈0.667,recall=TPTP+FN=23≈0.667\text{precision} = \frac{TP}{TP + FP} = \frac{2}{3} \approx 0.667, \quad \text{recall} = \frac{TP}{TP + FN} = \frac{2}{3} \approx 0.667

The false alarm rate is FP/(FP+TN)=1/3≈0.333FP / (FP + TN) = 1/3 \approx 0.333, and accuracy is 4/6≈0.6674/6 \approx 0.667. Dropping τ\tau to 0.4 would catch clip D and push recall to 1.0, at the price of still holding the one false alarm. Raising it to 0.6 would clear clip C and remove the false alarm, at the price of missing nothing new here but catching only the two obvious fakes. Small line moves, concrete price tags.

Do not use one score as sole proof for a takedown, a ban, or an accusation. The method assumes generator traces survive to the clip you test. Recompression, resizing, cropping and re-uploads smear those traces, and each newly retrained generator erases the cue you just published. When the assumption fails the score degrades gracefully into noise, confident and wrong, which is why high stakes calls pair detection with provenance and human review.

Code

This snippet replays the worked example so you can feel the threshold move. Change tau and watch the hits trade against the alarms.

def metrics_at_threshold(scores: list[float], labels: list[int], tau: float = 0.5) -> tuple[float, float, float]:    preds: list[int] = [1 if s >= tau else 0 for s in scores]    tp = sum(1 for p, y in zip(preds, labels) if p == 1 and y == 1)    fp = sum(1 for p, y in zip(preds, labels) if p == 1 and y == 0)    fn = sum(1 for p, y in zip(preds, labels) if p == 0 and y == 1)    tn = sum(1 for p, y in zip(preds, labels) if p == 0 and y == 0)    precision = tp / (tp + fp)    recall = tp / (tp + fn)    fpr = fp / (fp + tn)    return precision, recall, fpr

scores: list[float] = [0.92, 0.81, 0.55, 0.48, 0.30, 0.12]labels: list[int] = [1, 1, 0, 1, 0, 0]  # 1 means truly fakep, r, f = metrics_at_threshold(scores, labels, tau=0.5)print(f"{p:.3f} {r:.3f} {f:.3f}")  # -> 0.667 0.667 0.333

Those three numbers match the hand count above: precision 2/3, recall 2/3, false alarm rate 1/3. Try tau=0.4 and tau=0.6 to see the trade in your own terminal.

Watch Out For

Setting the threshold on the clips you report

You slide τ\tau until the demo set looks perfect, then quote those numbers. On fresh uploads the false alarms spike and the misses return, because the line memorized your demo instead of generalizing.

Symptom: sparkling metrics in the notebook, embarrassing ones in production, with the gap opening the week after launch. Fix: choose τ\tau on a held out split, freeze it, and report metrics once on data the threshold never saw. Retune on a schedule, since generator cues drift.

Reading one score as a verdict on laundered clips

A clip arrives recompressed, cropped, captioned and re-uploaded three times. The detector returns a confident low score and you clear it. Laundering smears exactly the traces the score feeds on, so low no longer means real.

Symptom: confident clears on viral copies of clips your system flagged at upload. Fix: treat the score as triage, not verdict. Corroborate with provenance, source context and a second look, and never rest a high stakes call on one number alone.

The Quick Version

  • A detector maps each clip to a fakeness score and calls it fake at or above a threshold you choose.
  • The score blends body and physics consistency with statistical traces like upsampling peaks and blending seams.
  • On the six worked clips at 0.5, precision and recall are both 0.667, with one false alarm and one miss.
  • Moving the threshold trades false alarms against misses. No position removes both errors.
  • Detection assumes generator traces survive, so pair it with provenance and human review on anything high stakes.