Beyond Scam Ads - NASK’s 24-Hour Challenge to Uncover Connected Campaigns

A suspicious advertisement is not always an isolated incident. NASK’s upcoming HackNation challenge focuses on fraudulent ads that change their wording, appearance and web addresses while remaining part of the same operation. The engineering problem is to connect those fragments into a campaign that investigators can examine.
HackNation takes place in Bydgoszcz, Poland, on November 21-22, 2026. NASK’s challenge has a particularly concrete destination: intended integration into DROZD, its developing deepfake-detection system. That makes the event more than an exercise in building an impressive demo, but it does not make a 24-hour prototype production-ready.
The useful question for development teams is therefore not whether they can finish a scam-detection platform overnight. It is whether they can demonstrate a credible, testable approach to campaign linking, with a clear path toward operational use.
What the Challenge Actually Requires
NASK’s “Kolekcjoner Fejków” challenge asks participants to analyze advertisement text, images, domains and other digital traces, then identify similarities that connect related materials into larger scam campaigns. The focus is the relationship between advertisements, not simply whether an individual image has been manipulated.
The organizer schedules publication of the full task statements for 11:00 on November 21, with coding ending at 11:00 on November 22. The public teaser should therefore not be treated as a complete technical specification. Dataset sizes, evaluation rules, integration interfaces and deployment acceptance criteria should not be assumed in advance.
A sound engineering interpretation separates three questions:
Manipulation: Has an image, video or other asset been altered?
Fraud: Does the advertisement support a deceptive offer or interaction?
Campaign membership: What evidence connects this material to other suspicious materials?
For the prototype, these should be separate outputs rather than one undifferentiated “scam score.” A campaign-linking result should explain a relationship, not imply that the system has established criminal ownership.
This broader scope also reflects the threat patterns described in NASK’s pre-holiday fraud guidance: copied logos, familiar colors, imitated page layouts, misleading offers and links followed by calls from impersonated consultants. Deepfakes are only one part of that picture.
Design an Evidence Pipeline, Not a Single AI Verdict
A practical prototype should prioritize traceable evidence over architectural complexity. The following is a proposed design, not a description of DROZD’s internal architecture.
Preserve observations before extracting features
Create an evidence record for each advertisement. Where available and permitted, retain:
The original image or screenshot.
Advertisement text and text extracted from the image.
The observed URL and capture timestamp.
Supplied landing-page or redirect information.
Source identifiers and processing status.
Keep original observations separate from derived features. Record missing information explicitly rather than treating an unavailable image or failed text extraction as evidence of dissimilarity.
This design would let an analyst inspect the material behind a proposed connection and let developers reproduce a result after changing the matching logic.
Compare several kinds of evidence
Start with a small, explicit set of matching methods. Candidate approaches to test include:
Text: Compare extracted wording, recurring offers and named entities.
Images: Test exact matching, perceptual fingerprints or pretrained visual representations.
Web traces: Compare observed domains, URL structures and supplied destination information.
Time: Examine when related materials were observed.
Treat these as hypotheses to validate against the available data. Do not assume that every modality improves the result.
For example, the prototype should not accept a shared logo as sufficient campaign evidence. Instead, test whether combining visual similarity with matching offer language and destination traces produces more defensible links. Mark links based on only one weak signal for review.
Make relationships inspectable
Represent advertisements as nodes and proposed relationships as edges. Each edge should retain its supporting evidence, feature values and scoring version.
Avoid making every chain of similarities an automatic campaign merge. If advertisement A resembles B, and B resembles C, require additional checks before presenting all three as one operation.
The interface should answer a simple question: Why are these advertisements grouped together? That explanation is more useful for investigation than an unexplained cluster identifier.
Scope the 24 Hours Around a Testable Demonstration
The recommended target is one complete workflow:
Ingest evidence, propose relationships, form candidate groups and let a reviewer inspect the reasons.
A suggested allocation would be:
Hours 0-4: Read the released specification, inspect the data and establish a baseline.
Hours 4-10: Implement ingestion and a limited set of text, image and URL features.
Hours 10-16: Add relationship scoring and grouping, keeping weak links visible.
Hours 16-21: Build an evidence-review interface and evaluate errors.
Hours 21-24: Stabilize the demonstration, document limitations and prepare reproducible execution instructions.
This is a planning framework, not the organizer’s prescribed workflow.
Keep the first baseline deliberately simple. Begin with exact matches and clearly defined overlap rules. Add more complex representations only where testing shows a benefit.
The demonstration should include difficult cases as well as successful ones. Test unrelated advertisements with similar branding, related advertisements with changed wording, missing modalities and uncertain relationships. Allow the system to return “insufficient evidence” instead of forcing every item into a campaign.
For a prospective institutional user, the prototype’s most important deliverable should be an inspectable evidence package: proposed links, explanations, evaluation results and documented failure cases.
Measure Campaign Quality Before Claiming Readiness
Evaluation should match the task. A proposed assessment plan should measure incorrect campaign merges separately from missed relationships.
Where reliable campaign labels exist, clustering metrics offer useful distinctions. Homogeneity measures whether clusters contain members of a single reference class. Completeness measures whether members of a reference class remain together. The adjusted Rand index provides a chance-adjusted comparison between predicted and reference groupings. These measures require suitable reference labels; they cannot establish campaign truth by themselves.
For this use case, combine those measures with operational questions:
How many proposed links can reviewers confirm?
Which legitimate-looking similarities cause incorrect merges?
How often does the system split a known campaign?
What happens when an image, domain trace or text field is missing?
How long does processing take, and how much review work does it create?
Design the test set to challenge the intended deployment. Keep exact duplicates from inflating the assessment, test changed creative assets separately, and use chronological evaluation where trustworthy timestamps permit it. Report uncertain or incomplete labels as a limitation.
Before deployment, also define security boundaries for evidence collection. Fetching arbitrary submitted URLs creates server-side request forgery risks, and automatic redirects can bypass initial validation. OWASP recommends application- and network-layer protections and disabling automatic redirect following in the relevant validation workflow.
For this detector, a conservative implementation plan should isolate collection from core services, restrict network access and avoid live fetching entirely during the prototype if supplied evidence is sufficient.
Treat DROZD Integration as a Separate Engineering Stage
NASK names DROZD as the intended integration destination. Its announcement describes an objective, not a completed integration or confirmed production deployment. The event also presents further development as an opportunity rather than an automatic outcome.
A credible handover should therefore identify the decisions still required:
Who owns maintenance and further development?
What evidence schema and interface will the receiving system accept?
What error rates and review workload are acceptable?
How will analysts correct links or split incorrectly merged groups?
What access controls, retention rules and audit records are required?
Who approves security testing, rollout and rollback?
The proposed deployment path should begin with a controlled pilot in which analysts review suggestions without triggering automatic enforcement. Progress should depend on measured usefulness, security testing and agreed acceptance criteria.
Twenty-four hours can be used to test an engineering hypothesis. Production adoption requires a separate commitment to integration, evaluation and operational ownership. For teams building campaign-detection tools, the strongest outcome is not the broadest AI claim. It is a system that makes its evidence clear enough to test, challenge and improve.



