Design Deepfake Detection - System Design | WinJob

Problem statement: synthetic media verification at scale

platforms ingest suspect video, audio, and images; return calibrated fake probability with audit trail for newsrooms and identity providers

Numbers to state early

Metric A: 50M scans/day
Metric B: p99 8s async
Metric C: 99.5% availability

Mechanism

The hot path Publisher → Ingest → Verdict API must preserve evidence chain-of-custody. For problem statement: synthetic media verification at scale, cite 50M scans/day when challenged on scale. Lead with calibrated fake probability, not binary labels, because downstream moderation policies differ per tenant.

Failure and edge cases

Codec mismatch after transcode, duplicate Idempotency-Key replays, GPU preemption mid-ensemble, tenant threshold misconfiguration causing review-queue floods, and C2PA manifest signature failures that should not block ML fallback.

When discussing Problem statement: synthetic media verification at scale, anchor on multimodal deepfake detection for Microsoft/Sensity-style platforms—not generic "ML API" boxes. Mention p99 8s async before naming GPU SKUs.

Design pressure specific to deepfake detection

Operators running Ingest under problem statement: synthetic media verification at scale should assume partial GPU regions: widen uncertainty bands, route borderline scores to human review, and never delete original blobs until legal hold expires. Tenants need 99.5% availability enforced at the API gateway.

Java

javaOne Dark Pro

1 public final class DetectionJobState {
2   public enum Status { QUEUED, TRANSCODING, INFERENCE, REVIEW, COMPLETED, FAILED }
3  
4   private final String jobId;
5   private final Status status;
6   private final double fakeScore;
7  
8   public DetectionJobState(String jobId, Status status, double fakeScore) {
9     this.jobId = jobId;
10     this.status = status;
11     this.fakeScore = fakeScore;
12   }
13  
14   public boolean needsHumanReview(double low, double high) {
15     return status == Status.INFERENCE && fakeScore >= low && fakeScore <= high;
16   }
17 }

Python

pythonOne Dark Pro

1 from dataclasses import dataclass
2 from enum import Enum
3  
4 class Status(str, Enum):
5     QUEUED = "queued"
6     TRANSCODING = "transcoding"
7     INFERENCE = "inference"
8     REVIEW = "review"
9     COMPLETED = "completed"
10     FAILED = "failed"
11  
12 @dataclass(frozen=True)
13 class DetectionJobState:
14     job_id: str
15     status: Status
16     fake_score: float
17  
18     def needs_human_review(self, low: float, high: float) -> bool:
19         return self.status == Status.INFERENCE and low <= self.fake_score <= high

TypeScript

typescriptOne Dark Pro

1 export type DetectionStatus =
2   | "queued"
3   | "transcoding"
4   | "inference"
5   | "review"
6   | "completed"
7   | "failed";
8  
9 export interface DetectionJobState {
10   jobId: string;
11   status: DetectionStatus;
12   fakeScore: number;
13 }
14  
15 export function needsHumanReview(
16   job: DetectionJobState,
17   low: number,
18   high: number,
19 ): boolean {
20   return job.status === "inference" && job.fakeScore >= low && job.fakeScore <= high;
21 }

Why interviewers care

Deepfake Detection interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem statement: synthetic media verification at scale that proves you understand real outages, not happy-path diagrams.

Key Highlights

•50M scans/day
•Publisher → Verdict API
•platforms ingest suspect video, audio, and images; return calibrated fake probability with audit trail for newsrooms and.

1	public final class DetectionJobState {
2	public enum Status { QUEUED, TRANSCODING, INFERENCE, REVIEW, COMPLETED, FAILED }
3
4	private final String jobId;
5	private final Status status;
6	private final double fakeScore;
7
8	public DetectionJobState(String jobId, Status status, double fakeScore) {
9	this.jobId = jobId;
10	this.status = status;
11	this.fakeScore = fakeScore;
12	}
13
14	public boolean needsHumanReview(double low, double high) {
15	return status == Status.INFERENCE && fakeScore >= low && fakeScore <= high;
16	}
17	}

1	from dataclasses import dataclass
2	from enum import Enum
3
4	class Status(str, Enum):
5	QUEUED = "queued"
6	TRANSCODING = "transcoding"
7	INFERENCE = "inference"
8	REVIEW = "review"
9	COMPLETED = "completed"
10	FAILED = "failed"
11
12	@dataclass(frozen=True)
13	class DetectionJobState:
14	job_id: str
15	status: Status
16	fake_score: float
17
18	def needs_human_review(self, low: float, high: float) -> bool:
19	return self.status == Status.INFERENCE and low <= self.fake_score <= high

1	export type DetectionStatus =
2	\| "queued"
3	\| "transcoding"
4	\| "inference"
5	\| "review"
6	\| "completed"
7	\| "failed";
8
9	export interface DetectionJobState {
10	jobId: string;
11	status: DetectionStatus;
12	fakeScore: number;
13	}
14
15	export function needsHumanReview(
16	job: DetectionJobState,
17	low: number,
18	high: number,
19	): boolean {
20	return job.status === "inference" && job.fakeScore >= low && job.fakeScore <= high;
21	}