World Congress 2026 Europe • Jul 10, 2026 • Session details

PET ARENA Season 2.

Sagar Sharma .

Static benchmarks leave privacy systems vulnerable. Now, TikTok and Oblivious have launched a live, peer-to-peer battleground where defenders and attackers co-evolve to battle-test differential privacy architectures.

Pause
Mute Enter Fullscreen
#1 about 2 min

Replacing static privacy benchmarks with adaptive adversarial testing

Peer-to-peer competitions reveal how privacy systems fail against evolving query probes.

#2 about 2 min

Fostering a community of privacy builders and breakers

Data engineers and researchers compete to identify empirical privacy leakages using basic Python and SQL.

#3 about 2 min

Analyzing statistical and logical differencing attack strategies

Prior attackers successfully exploited systems via subgroup inference and reverse-engineered noise mechanisms.

#4 about 2 min

Avoiding common privacy querying and budget management pitfalls

Failing to account for noise floors or exhausting privacy budgets leads to refused queries and lost points.

#5 about 2 min

System architecture orchestrating automated privacy match evaluations

A centralized evaluation hub manages datasets, API gateways, and query proxies to prevent overfitting.

#6 about 1 min

Implementing defense track noise mechanisms and query guardrails

Defenders must balance query suppression and custom noise mechanisms without failing hidden utility thresholds.

#7 about 2 min

Designing inference and record linkage attacks for red teaming

Attackers attempt to extract sensitive attributes and bypass dynamic defenses to reidentify database targets.

#8 about 2 min

Progressing through peer-to-peer cycles and final calibration rounds

Participants complete initial house missions before facing opponent submissions across three competitive cycles.

#9 about 2 min

Calculating leaderboard scores with a dynamic rating system

Continuous match scores scale zero-sum algorithms and adapt K factors based on track participation density.

#10 about 2 min

Preventing match collusion and competing for dual-track rewards

System guardrails involuntarily pair opponents and enforce query limits to ensure equitable grandmaster evaluations.

#11 about 5 min

Building and testing strategies in the browser sandbox environment

A user journey demonstrating how to configure LLM classifiers and execute budgeted attacker notebooks within the interface.

#12 about 3 min

Co-evolving submissions and managing late competition registration limits

Continuous strategy modifications reward participants who acclimate to evolving opponent models early in the cycle.

Matching moments

1:51 min

Exploring inference, extraction, and the adversarial threat landscape

Nura Kawa · World Congress 2023

17:05 min

Navigating data privacy boundaries and adversarial model reliability

Alexandra Waldherr · LIVE

4:04 min

Embedding data security and applied ethics into developer education

Daniel Tao +3 · World Congress 2024

4:17 min

Overcoming commercial constraints and exploring transparent multiplayer artificial intelligence

Johanna Pirker Johanna Pirker · World Congress 2021

1:32 min

Developing fail-safes and defenses through adversarial training

Nura Kawa · World Congress 2023

5:55 min

Iterating extensively through evaluations and red teaming

Seppe Housen Seppe Housen · Europe 2026 Virtual