PaperlayerView exact revision on arXiv

Source-linked AI summary

Same-Player Verification for Account Consistency in Counter-Strike 2

Xuchen Zhang

arXiv:2608.24893v1cs.AIcs.LG

TL;DR

Account-history consistency review needs a scalable way to compare a current CS2 operator with behavior across multiple historical matches. This paper formulates the task as open-set same-player verification using behavioral fingerprints, achieving 0.931 average ROC AUC and 0.986 AUC at K=10 history aggregation.

  • Problem

    Account-history consistency review requires systematic comparison of current behavior with multiple historical matches, beyond existing admission, reporting, and manual-review workflows.

  • Method

    The paper learns an open-set pairwise same-player verification model from CS2 demo-player behavioral fingerprints spanning mechanical, economic, combat, and temporal behavior.

  • Results

    0.931 average ROC AUC and 0.722 different-player recall at 95% precision were achieved, while history aggregation increased AUC from 0.931 at K=1 to 0.986 at K=10.

  • Takeaways & Limitations

    CS2 demo behavior supports supervised same-player verification and account-level identity-consistency modeling, with strongest signals in low-level operations such as crosshair control and firing rhythm.

  • Takeaways & Limitations

    Noisy or contaminated identity labels mean deployment requires human confirmation, label audits, and feedback rather than treating one-time dataset labels as permanent truth.

Abstract

from arXiv · show

In competitive first-person shooter (FPS) games such as Counter-Strike 2 (CS2), account-integrity review often asks whether an account's recent behavior remains consistent with its historical operator. This consistency question arises in cases such as temporary substitution, rank boosting, and high-skill players using lower-ranked accounts, where manual review requires comparing a current match against multiple historical matches. We formulate this review task as same-player verification: we encode the behavioral trajectory of a single player in a match replay (demo) as a demo-player behavioral fingerprint, and train a model to judge whether two behavioral observations come from the same real player. Grounded in CS2 game understanding, the fingerprints cover crosshair control, movement-stop-fire coordination, economy/buy, combat/engagement, and temporal rhythm. From 1,330 CS2 demos we extract 13,300 demo-player observations, and sample 663,590 same/different pairs from an 88.4M candidate-pair space for supervised training and evaluation. The final pairwise model reaches an average ROC AUC of 0.931 and achieves 0.722 different-player recall at 95% precision. Feature analysis shows that the strongest identity signals come from low-level operations, especially crosshair control, firing rhythm, and movement-stop-fire coordination, indicating that stable low-level mechanical habits are more informative for this verification task than single-match performance outcomes. In the account-history aggregation evaluation, increasing history depth raises AUC from the K=1 single-pair baseline of 0.931 to 0.986 at K=10. These results show that CS2 demo behavior can support supervised same-player verification and account-level identity-consistency modeling through multi-demo history aggregation.

I. INTRODUCTION … D. Contributions

The paper frames account-history consistency in CS2 as open-set same-player verification using behavioral fingerprints that compare current and historical demo-player behavior. It combines low-level behavioral signals with sequence representations and multi-demo aggregation, achieving strong pairwise and account-level verification results.

  • A. Problem Background: Account sharing, rank boosting, temporary substitution, and lower-ranked-account use can make an account’s current operator inconsistent with historical behavior, threatening matchmaking fairness and trust.
  • A. Problem Background: Manual demo review and player reports do not systematically compare current behavior against multiple historical matches, motivating an automated account-history consistency method.
  • B. Paper Idea: The method models account-history review as open-set same-player verification, learning whether two demo-player behavior segments come from the same real player without identifying that player.
  • B. Paper Idea: Fingerprints emphasize lower-level operations and timing, including crosshair control, firing rhythm, movement-stop-fire coordination, economy, combat, and temporal behavior.
  • B. Paper Idea: The pairwise model combines aggregate behavioral features with Transformer-derived sequence embeddings, while history aggregation converts current-versus-history scores into account-level consistency signals.
  • C. Results Overview: 1,330 CS2 demos yielded 13,300 demo-player observations and 663,590 sampled same/different pairs from an 88.4M candidate-pair space.
  • C. Results Overview: ROC AUC 0.931 and 0.722 different-player recall at 95% precision were achieved by the final pairwise model, with strongest identity signals in crosshair and combat micro-operations.
  • D. Contributions: The contributions establish a supervised verification formulation, identify low-level behavioral identity signals, and extend verification from single pairs to multi-demo account-history consistency modeling.

II. RELATED WORK … III. METHOD

The paper situates CS2 same-player verification within account-integrity review, Counter-Strike behavioral identity modeling, and broader behavioral biometrics. It addresses the gap between existing review mechanisms and systematic current-versus-history comparison by using stable CS2 habits and manually confirmed histories.

  • A. Account Identity Verification and Fairness Mechanisms on Competitive Platforms: Competitive-platform research identifies operator inconsistency and multi-account use as threats to matchmaking fairness and tournament credibility, including boosting-related skill discrepancies.Boosting is described as using a stronger player on an account to improve its rank or results.
  • A. Account Identity Verification and Fairness Mechanisms on Competitive Platforms: Existing platform mechanisms combine anti-cheat, identity verification, reporting, and manual review but lack systematic comparison of current behavior with multiple historical matches.This paper formalizes that comparison as CS2 demo-based same-player verification.
  • B. Counter-Strike and FPS Player Behavioral Identity Modeling: Prior Counter-Strike and CS2 studies use behavioral features to identify known professionals or distinguish known player pairs, with signals in aiming, shooting, movement state, and game context.Related work therefore establishes that CS/CS2 demo telemetry contains identity-related information.
  • III. METHOD: This study instead tests whether a current demo matches an account’s historical demos when the operator may be unknown or unseen during training.It focuses on crosshair control, firing rhythm, movement-stop-fire coordination, state switching, buying rhythm, and risk preference, using unseen-player splits and current-versus-history aggregation.
  • B. Counter-Strike and FPS Player Behavioral Identity Modeling: FPS fair-play methods flag skill-rating discrepancies or passive aimbots by combining mouse, keyboard, in-game-action, expert-knowledge, shooting-performance, and broader-skill evidence.These approaches target operator ability mismatches with account rank or broader skillfulness rather than direct same-player history consistency.
  • C. Broader Game Behavior and Behavioral Biometrics: Replay and telemetry research beyond FPS games models identity or style using mouse, game statistics, strategy, build order, unit control, resource management, and operation rhythm.Dota 2 work directly targets whether two matches were completed by the same player.
  • C. Broader Game Behavior and Behavioral Biometrics: Behavioral biometrics research finds identity signals in mouse trajectories, touchscreen habits, interaction rhythms, motion patterns, VR motion, mouse dynamics, and active-user behavior under shared accounts.These findings provide background support while differing from natural CS2 account-history review.
  • D. Positioning of This Paper: The paper combines platform review entry points, Counter-Strike identity signals, and broader behavioral-biometrics evidence into CS2 account-history consistency review.Its positioning uses manually confirmed account histories to construct pairwise labels and learn same-player verification.

A. Problem Definition · B. Game-Understanding-Based Behavioral Fingerprint Features

The paper defines same-player verification over pairs of CS2 demo-player observations and extends it to account-history aggregation. It represents each observation with a game-understanding-based behavioral fingerprint that captures operational and decision habits rather than only match outcomes.

  • A. Problem Definition: A demo-player observation is the structured behavioral record left by one player in one CS2 match demo.CS2 demos contain positions, view angles, weapons, events, and round states; a standard match usually yields 10 observations.
  • A. Problem Definition: Pairwise verification compares two demo-player behavioral trajectories and outputs an identity-consistency score for same-player likelihood.Higher scores indicate stronger evidence that the observations come from the same real player.
  • A. Problem Definition: Account-history aggregation computes K pairwise scores between a current observation and K historical observations, then combines them into an account-level consistency signal.The two-level formulation separates pairwise verification from account-history consistency assessment.
  • A. Problem Definition: The complete behavioral fingerprint concatenates the aggregate behavioral fingerprint with the combat-window sequence representation.Player labels supervise training and evaluation sample construction but are not required to identify the real player at test time.
  • A. Problem Definition: The pairwise model receives endpoint fingerprints, explicit comparison features, and pair-level context, with labels y_ij = 1 for same-player and y_ij = 0 for different-player pairs.Pair-level context encodes map relation only, excluding demo, match, teammate, and opponent identities.
  • B. Game-Understanding-Based Behavioral Fingerprint Features: The game-understanding fingerprint summarizes each player’s operation and decision habits from one CS2 match.Recovered signals include within-round position, view angle, movement, weapon state, firing, damage, utility, and buy events.
  • B. Game-Understanding-Based Behavioral Fingerprint Features: 252 per-demo-player behavioral fingerprint features describe frequencies, intervals, distribution shapes, and conditional relations across eight sub-representations.The features characterize movement, aiming, firing, state switching, equipment purchases, and utility use across maps and round phases.
  • B. Game-Understanding-Based Behavioral Fingerprint Features: Mechanics/state features capture micro-transition habits before firing, such as completely stopping or firing early while moving.Aggregating these counter-strafing-related transitions over an observation expresses a player’s long-formed operation style rather than duel outcomes.

C. Sequential Behavioral Representation

The method represents player identity through short combat-centered action sequences, addressing cases where similar aggregate behavior can arise from different round developments. Fixed-length event windows are encoded and aggregated into a demo-player-level sequence embedding.

  • Motivation: Similar firing, movement, or utility counts can reflect different movement-stop-fire coordination and therefore may not distinguish player behavior reliably.The motivating contrast is counter-strafing before micro-adjusting and firing versus firing while moving before counter-strafing and micro-adjusting.
  • Event-Centered Windows: Local windows are centered on firing, damage, kills, and death events, where identity-relevant behavior is concentrated.Death-related evidence is concentrated before the death event, while weak-signal portions of demos are excluded.
  • Event-Centered Windows: Each combat window is represented as a 32×16 continuous numerical tensor.The windows are fixed-length tick sequences cropped from demo-player observations around combat events.
  • Sequence Encoding: A sequence encoder maps each combat window to an embedding, and an aggregation function combines the window embeddings into a demo-player-level sequence embedding.For observation i, the number of available combat windows is denoted M_i.
  • Sequence Encoding: During training, each observation samples at most 24 windows, and the sequence encoder is implemented as a Transformer encoder.The sampling limit controls how many available combat windows contribute to an observation during training.

D. Pairwise Comparison Representation · E. Account-History Aggregation

The method augments paired demo-player fingerprints with explicit symmetric comparison features, then aggregates current-versus-history pairwise scores into account-level consistency measures. History aggregation supports both raw-score averaging and an interpretable evidence-scale transformation.

  • D. Pairwise Comparison Representation: Pairwise comparison adds symmetric features that explicitly encode absolute and relative differences between two demo-player fingerprints.This reduces the need for the model to infer feature differences solely from concatenated endpoint values.
  • D. Pairwise Comparison Representation: Given two raw fingerprints, the method constructs an explicit comparison block for each scalar behavioral feature.The comparison block is denoted compare(x_i,x_j).
  • D. Pairwise Comparison Representation: Numerical stabilization uses ε, while an indicator function and exclusive OR encode discrete comparison conditions.The indicator equals 1 when its condition holds and 0 otherwise.
  • E. Account-History Aggregation: For an account check, the current observation is compared with each of K historical observations, producing K pairwise consistency scores.The historical set is H = {x_h1,...,x_hK}.
  • E. Account-History Aggregation: Raw-score aggregation converts the K pairwise scores into an account-level consistency score, including the raw-score mean.This is described as the most direct aggregation approach.
  • E. Account-History Aggregation: An empirical LLR-style transformation maps each pair score to an evidence value before combining scores across history.The mapping uses same-player and different-player training-side score distributions, discrete bins, and a smoothing term α; positive evidence favors same-player pairs, while negative evidence favors different-player pairs.
  • E. Account-History Aggregation: The transformed evidence values are averaged over the K historical observations to obtain the account-level score.The evidence scale reflects how common each score bin is in the training-side same-player distribution relative to the different-player distribution.

F. Overall Workflow · IV. EXPERIMENTS · A. Dataset and Evaluation Setup

The workflow encodes demo-player observations as fingerprints, compares current and historical observations pairwise, and aggregates scores into account-level consistency evidence. Experiments use manually confirmed and professional-player CS2 demos, person-disjoint evaluation, sampled pair distributions, and dedicated views for history, time, and map effects.

  • F. Overall Workflow: At inference, fingerprints are compared pairwise and multiple current-versus-history scores are aggregated into an account-level consistency score.Training mainly supervises the pairwise comparator gθ.
  • A. Dataset and Evaluation Setup: 1,330 demos and 13,300 demo-player observations form the main dataset from active university CS players spanning 12 amateur competitive tiers.The collection entry point is a university CS player guild on Perfect World Arena.
  • A. Dataset and Evaluation Setup: 227 professional demos and 2,270 professional demo-player observations extend the dataset, including 631 target observations corresponding to 50 professional players.Professional identities come from public tournament records and provide externally verifiable identity sequences and higher skill coverage.
  • A. Dataset and Evaluation Setup: 66,359 same-player pairs are formed from 2,380 manually confirmed observations representing 64 confirmed persons.Confirmed observations are used to construct same-player positives by enumerating per-person combinations.
  • A. Dataset and Evaluation Setup: 663,590 pairwise samples are obtained from 88.4M possible pairs by retaining all same-player pairs and sampling different-player pairs at 1:9.Precision, AP, and recall at fixed precision are reported on this sampled evaluation distribution, so deployment thresholds require platform-specific recalibration.
  • A. Dataset and Evaluation Setup: Train/test splitting by person id prevents the same real player from appearing on both sides and removes crossing pairs.This protocol is designed to prevent identity leakage.
  • IV. EXPERIMENTS: The experiments evaluate pairwise behavior, professional-demo augmentation, history aggregation, time effects, and cross-map robustness.The main pairwise conclusions use the six-split average of T (1:6)pair, while Tpro, Thist(K), Ttime, and Tmap target the additional evaluation questions.
  • A. Dataset and Evaluation Setup: AP summarizes discrete precision-recall area, while recall at 95% precision measures recovered true inconsistent pairs or account-history groups at a high-confidence threshold.Map variation matters because maps affect routes, combat distances, and utility combinations, although stable player habits may persist across maps.

B. Behavioral Representations and Model Selection · 1) Representation Levels: Which Information Brings Gains:

Behavioral representations substantially outperform performance-only features for same-player verification, while combining behavioral fingerprints with action sequences and explicit pairwise comparisons yields further gains. The strongest improvement comes from explicit pairwise comparison features, which directly encode absolute and relative behavioral differences.

  • B. Behavioral Representations and Model Selection: Table III reports the main pairwise result using the average over T (1:6).The same LightGBM pairwise model compares the consistency signal provided by different input representations on each split.
  • 1) Representation Levels: Which Information Brings Gains:: E1’s outcome/performance-only representation is a weak baseline at AUC 0.599.It uses result-based statistics including K/D, damage, score, headshot rate, and first kill / first death.
  • 1) Representation Levels: Which Information Brings Gains:: E2 raises AUC to 0.831 by using behavioral dimensions closer to player identity than performance outcomes.These include crosshair control, movement-stop-fire coordination, combat rhythm, economy/buy, and round timing.
  • 1) Representation Levels: Which Information Brings Gains:: E3 reaches AUC 0.723 using sequence representation alone, showing that action order in combat windows carries identity information.Its information density is lower than the behavioral representation.
  • 1) Representation Levels: Which Information Brings Gains:: E4 improves AUC by 0.011 on average and @95% precision recall by 0.053 on average over E2.The behavioral fingerprints and sequence embeddings are complementary: one summarizes recurring operation and decision distributions, while the other preserves action order.
  • 1) Representation Levels: Which Information Brings Gains:: E5 improves AUC by 0.089 on average and @95% precision recall by 0.310 on average over E4.Explicit comparison features directly use absolute and relative differences on the same behavioral dimensions.
  • B. Behavioral Representations and Model Selection: Explicit comparison features reduce the burden on the model to learn symmetric difference relations when samples are limited and behavioral features are heterogeneous.This is the stated motivation for using explicit pairwise comparisons in the model input.

2) Model Selection and High-Precision Inconsistency Retrieval:

Model selection favored a continuous numerical combat-window Transformer for sequence features and LightGBM for pairwise scoring. On the analysis split, LightGBM achieved best F1 0.855 and recalled about 0.816 of different-player pairs at 95% precision.

  • Sequence feature model selection: The continuous numerical combat-window Transformer was the most stable sequence-feature model among the compared representations and aggregation/readout methods.It was trained with player identity labels and models action order around combat windows.
  • Pairwise model selection: LightGBM was strongest on T (1) pair, with XGBoost close behind, while fast MLP was clearly weaker.Tree-based models more stably used continuous statistics, sparse indicators, map information, and explicit differences at the current data scale.
  • Pairwise model selection: The final implementation used game-understanding-based behavioral fingerprints, a Transformer-derived sequence representation, and a LightGBM pairwise scorer.The rank-average ensemble did not exceed the strongest single model.
  • High-precision inconsistency retrieval: 0.855 best F1 was reached by the LightGBM final representation on T (1) pair.The result comes from the different-player retrieval precision-recall evaluation.
  • High-precision inconsistency retrieval: 0.816 of different-player pairs were recalled at 95% precision on T (1) pair.This threshold region supports retrieval under a low false-positive budget for account-history consistency review.

C. Feature-Family Sensitivity: Which Behaviors Carry Identity Signal · D. Account-History Aggregation: From Pairwise Scores to Multi-Demo History Comparison

Feature sensitivity identifies aiming/crosshair, mechanics/state, and combat/engagement as the main identity-bearing behaviors, with other families contributing supplementary information. Aggregating pairwise evidence across historical demos substantially improves account-level verification over a single comparison.

  • C. Feature-Family Sensitivity: Which Behaviors Carry Identity Signal: Aiming/crosshair is the strongest identity signal, with full xbeh reaching 0.929 average AUC and removal reducing AUC by 0.074 on average.The six removal drops are -0.099/-0.059/-0.061/-0.084/-0.067/-0.073.
  • C. Feature-Family Sensitivity: Which Behaviors Carry Identity Signal: Aiming/crosshair alone reaches 0.883 AUC, only 0.047 below the full representation and stronger than other single families.Mechanics/state forms the second tier among individual feature families.
  • C. Feature-Family Sensitivity: Which Behaviors Carry Identity Signal: Using aiming/crosshair, mechanics/state, and combat/engagement lowers AUC by only 0.007, whereas removing all three lowers AUC by 0.175.Rhythm, space, buying, and context preferences provide supplementary information.
  • D. Account-History Aggregation: From Pairwise Scores to Multi-Demo History Comparison: Account-history aggregation compares one current observation with K historical observations within the held-out side of a person-disjoint split.The mean LLR aggregates the K pairwise scores into an account-level consistency score.
  • D. Account-History Aggregation: From Pairwise Scores to Multi-Demo History Comparison: Mean LLR AUC rises from 0.931 at K=1 to 0.971 at K=3 and 0.986 at K=10.Multiple historical demos provide a more stable behavioral reference than a single pairwise score.
  • D. Account-History Aggregation: From Pairwise Scores to Multi-Demo History Comparison: The aggregation gain comes mainly from accumulating multiple pairwise evidence, since raw-score mean gives very similar AUC.Single matches can shift with map, teammates, weapons, opponents, or the player’s state that day.

E. Cross-Time, Cross-Map, and Data Expansion Experiments … 3) Training Expansion Experiment with Public Professional Demos:

Cross-time and cross-map comparisons remain effective but are harder than baseline pairings, while adding public professional demos does not consistently improve model performance. The experiments show that recognition remains usable under realistic history-review conditions despite temporal and map variation.

  • 1) Time Gap and Cross-Window Evaluation:: Time-separated comparisons are harder than random pairs, yet the model retains strong cross-time recognition under the high-precision threshold.Longer gaps can change player state, map pool, version, settings, and play style; the time condition has an AUC about 0.008 lower than the pair condition.
  • 1) Time Gap and Cross-Window Evaluation:: 18.96 days was the median time gap, with p90 at 24.68 days and a maximum of 31.07 days.These gaps reflect realistic comparisons between recent suspicious demos and earlier account-history demos.
  • 2) Cross-Map Robustness:: Same-map comparisons provide stronger evidence, with same-map AUC about 0.020 higher than T (1) pair and cross-map AUC about 0.008 lower.Cross-map comparisons are closer to real account-history review because same-map samples are usually fewer and longer spans increase difficulty.
  • 2) Cross-Map Robustness:: Cross-map recognition remains usable despite being harder than same-map comparison.Same-map pairs usually have higher recognition accuracy, whereas cross-map pairs better approximate realistic account-history review.
  • 3) Training Expansion Experiment with Public Professional Demos:: The professional-demo expansion experiment tested whether public professional match identity sequences in Tpro could improve training coverage and model performance.The reported results instead indicate that performance gains were not stable across models or expansion strategies.
  • 3) Training Expansion Experiment with Public Professional Demos:: Adding professional positives and negatives changed LightGBM AUC / @95% precision recall by -0.005 / -0.019 and XGBoost by 0.000 / +0.005.These results do not support a stable improvement from adding public professional demos to the main model.
  • 3) Training Expansion Experiment with Public Professional Demos:: Adding only professional positives changed LightGBM AUC / @95% precision recall by -0.002 / 0.000 and XGBoost by +0.002 / +0.008.The positive-only expansion produced a small XGBoost improvement while LightGBM was essentially flat.

V. DISCUSSION AND LIMITATIONS … VI. CONCLUSION AND FUTURE WORK

The paper formalizes account-history consistency review as open-set same-player verification using CS2 demo-player behavioral fingerprints and pairwise learning. Results support accurate verification and lightweight post-extraction scoring, while evaluation, label, deployment, and scalability boundaries motivate broader future validation.

  • A. Evaluation Boundaries: The evaluation is person-disjoint and observation-disjoint, but not fully match-disjoint, because different players from one demo may appear on opposite sides.Inputs exclude demo IDs, match IDs, teammate/opponent identities, and shared-match identifiers.
  • A. Evaluation Boundaries: Removing same-demo pairs changes AUC from 0.955 to 0.952 and AP from 0.936 to 0.937.This sanity check addresses same-demo co-occurrence rather than full match-disjoint evaluation.
  • A. Evaluation Boundaries: Player-clustered bootstrap produces wider uncertainty intervals but does not change the core conclusions.Pairwise construction can give players with more observations disproportionate influence and make pair-level uncertainty too narrow.
  • B. Data and Label Boundaries: Labels may be noisy because identity mappings can miss shared operators, while account sharing, multi-account use, and temporary substitution can contaminate both label types.Same-player labels mainly use manually confirmed account histories and public professional identities; different-player labels use identity mappings and negative-sampling rules.
  • C. Practical Deployment: Historical Reference Quality, Runtime Cost, and Responsible Use: History aggregation requires historical demos to mainly represent one real player; mixed histories should undergo a history self-consistency audit before deployment.The audit compares history-history pairwise scores to judge whether the history set is suitable as a reference.
  • C. Practical Deployment: Historical Reference Quality, Runtime Cost, and Responsible Use: About 24k pairs per second are scored on Apple M4 once fingerprints or sequence embeddings are available.Demo parsing and feature extraction remain the main offline costs, measured at seconds per demo on a small representative subset, with throughput depending on demo length, parallelism, and I/O.
  • VI. CONCLUSION AND FUTURE WORK: The paper formalizes account-history consistency review as open-set same-player verification using per-demo-player behavioral fingerprints and pairwise same-player consistency learning.The formulation targets whether a current match remains behaviorally consistent with an account history.
  • VI. CONCLUSION AND FUTURE WORK: 0.931 average ROC AUC and 0.722 different-player recall at 95% precision demonstrate accurate pairwise verification after fingerprint extraction.Feature analysis identifies crosshair control, firing rhythm, and movement-stop-fire coordination as especially important low-level identity signals.
Loading 2608.24893v1…