PaperlayerView exact revision on arXiv

Source-linked AI summary

PlayerEmbed: Learning Behavioral Fingerprint Representations from Raw Input Sequences

Tobias Schneider, Lorenz Sparrenberg, Priya Tomar, Shahzeb Qamar, Tobias Deußer, Rafet Sifa

arXiv:2610.32727v1cs.LGcs.AI

TL;DR

Raw game inputs contain fine-grained behavioral patterns, but noise and complexity make large-scale analysis difficult. PlayerEmbed learns compact embeddings from raw sequences and evaluates player identification on held-out samples and identities unseen during training. The reported findings show strong identification for known and unknown players, with larger training-player sets improving generalization to unseen identities.

  • Problem

    Raw player inputs contain behavioral information, but noise and complexity make useful patterns challenging to extract and large-scale manual analysis impractical.

    Source: §I · p. 1

  • Method

    PlayerEmbed uses deep metric learning to create compact, discriminative embeddings directly from raw player input sequences without manual feature engineering.

    Source: §I · p. 1

  • Results

    PlayerEmbed achieves strong identification on 1,000 known players and 1,000 additional players unseen during training, and increasing the number of known training players systematically improves generalization to unseen players.

    Source: §I · p. 2, §V · p. 7

  • Takeaways & Limitations

    Transfer to unseen identities and the observed cluster structure suggest the embeddings capture behavioral patterns beyond player identity and may support other downstream tasks.

    Source: §V · p. 7

    Further work could scale beyond the current 1,000 known and 1,000 unknown players using larger, more diverse datasets across games.

    Source: §V · p. 7

Abstract

from arXiv · show

Raw player input sequences in video games contain fine-grained behavior, but are high-dimensional and noisy, which makes large-scale analytics challenging. To address this, we introduce PlayerEmbed, a framework that learns compact and discriminative player embeddings using deep metric learning. Our approach uses a hybrid CNN-LSTM encoder trained on raw input sequences from 1,000 players. We evaluate the model on held-out sequences from these players and on sequences from 1,000 players entirely unseen during training to assess generalization to unseen identities. For player identification, a nearest-centroid classifier built from reference embeddings achieves strong performance, reaching balanced accuracies of 0.950 for players seen during training and 0.879 for players unseen during training. Systematically increasing the number of training players substantially improves generalization to unseen players, indicating a transition from simple fingerprinting to learning transferable behavioral patterns.

I. INTRODUCTION

Raw game inputs contain behavioral information, but noise and complexity make useful patterns difficult to extract and manual analysis hard to scale. PlayerEmbed learns compact, discriminative embeddings from raw sequences and evaluates player identification at scale, including on players unseen during training.

I. INTRODUCTION

  • Compact representations of noisy, complex inputs could support downstream behavior analysis while reducing manual, game-specific preprocessing.The paper notes that manual analysis is time-consuming, limited in scope, and increasingly impractical for games with many players.Source: §I · p. 1
  • Player identification asks which individual produced an input sequence and is useful for monitoring account integrity.Examples include account sharing, account hijacking, and smurfing, which can undermine matchmaking integrity.Source: §I · p. 1
  • PlayerEmbed learns compact, discriminative embeddings from raw player inputs using deep metric learning, without manual feature engineering.The framework focuses on keystroke and mouse dynamics, which are described as mostly game-agnostic.Source: §I · p. 1
  • DDNet offers diverse player skill and dynamic movement despite a relatively simple input schema, making players non-trivial to distinguish.Its open-source servers also enable player-input data collection.Source: §I · p. 1
  • The evaluation uses up to 1,000 training players and a further 1,000 players unseen during training, testing generalization to new identities.The authors describe this as a scale substantially larger than prior work.Source: §I · p. 2

II. RELATED WORK

Prior game-behavior research commonly relies on aggregated features or observations of gameplay rather than raw player inputs. PlayerEmbed addresses the comparatively unexplored task of learning representations directly from raw input sequences and evaluates identification on a larger pool including unseen players.

II. RELATED WORK

  • Gameplay representation studies often use images, video, or audio rather than the player’s underlying interaction signals –.Such approaches have been applied to genre recognition and player-experience detection; DeepLogger attempts to reconstruct inputs from video.Source: §II · p. 2
  • Representation learning directly from raw game input sequences remains largely unexplored compared with approaches using preprocessed behavior data.Prior formats include discrete-action logs,, aggregated gameplay statistics –, and provenance graphs.Source: §II · p. 2
  • Behavioral biometrics often use keystroke and mouse patterns, but many studies evaluate fewer than 100 people and assume all users are known during training,,.Player-identification work likewise often uses hand-crafted features and small, closed player pools,, [25].Source: §II · p. 2
  • Proxy-based metric learning reduces pairwise comparison burden during training, and Proxy Anchor Loss models data-to-data relations while retaining training efficiency.Triplet Margin Loss can train slowly because of its computational complexity,.Source: §II · p. 2
  • Many game-behavior studies use aggregated features, while DDNet churn prediction also combines aggregated behavior with machine learning and large language models.Earlier work on player behavior analytics has likewise relied on preprocessed aggregated features,,.Source: §II · p. 2

A. Data Collection & Preprocessing

The study collects anonymized DDNet input sequences and standardizes fixed-length samples for analysis. Its data selection includes a cap on per-player samples, and the framework overview depicts their use in nearest-centroid player identification.

A. Data Collection & Preprocessing

  • DDNet servers record received player inputs at each physics-engine update, which runs at 50 timesteps per second.Source: A. Data Collection & Preprocessing · p. 2
  • Each timestep uses eight features: movement and button states, cursor angle and distance, and horizontal and vertical player velocity.Continuous features are individually Z-score standardized using training-data statistics.Source: A. Data Collection & Preprocessing · p. 2
  • The collected sequences cover 950 maps and are linked only to randomly assigned player IDs to preserve privacy.The inputs were recorded anonymously from real players.Source: A. Data Collection & Preprocessing · p. 2

C. Train Model

  • Figure 1 presents the PlayerEmbed workflow for player identification using nearest-centroid classification.Source: Figure 1 · p. 3
  • Each sample represents 1,000 timesteps, or 20 seconds of behavior, and players contribute 53 to 2,585 available sequences before a 1,000-sequence cap.The cap retains 75% of total input sequences while limiting the top 235 players to up to 5.56 hours each.Source: A. Data Collection & Preprocessing · p. 2, C. Train Model · p. 3

B. Dataset Construction

The evaluation separates known players into training, validation, and held-out test samples, while reserving separate identities as unknown validation and test sets. A hybrid CNN-LSTM converts input sequences into normalized 128-dimensional embeddings; its architecture was selected using a separate dataset.

B. Dataset Construction

  • Known-player samples are split approximately 56% for training, 19% for validation, and 25% for test, with whole sessions held out from training.The next 1,000 identities form the unknown test set, and the subsequent 1,000 form unknown validation; those identities are unseen during training.Source: B. Dataset Construction · p. 3

C. Model

  • The CNN extracts local temporal and cross-feature patterns before the LSTM models longer-range dependencies.Depthwise separable convolutions reduce computational cost relative to standard convolutions; the convolution uses a kernel size of five and 128 output channels.Source: C. Model · p. 3
  • The final LSTM hidden state is projected and L2-normalized into a 128-dimensional embedding.The single-layer unidirectional LSTM has hidden size 128, and the projection head uses a hidden dimension of 256.Source: C. Model · p. 4
  • Architecture and hyperparameters were chosen on a separate, smaller dataset rather than tuned on the final dataset.The authors state that further tuning on validation subsets might improve performance.Source: C. Model · p. 4

D. Training

PlayerEmbed is trained with Proxy Anchor Loss on normalized embeddings, using uniform sequence sampling and validation-based model selection. Player identification compares similarity-based classifiers, reference/query splits, and query-embedding averaging.

D. Training

  • Proxy Anchor Loss trains L2-normalized embeddings with cosine similarity; uniform sampling is used because class-weighted sampling and MPer-Class batching did not measurably improve performance.The loss converged substantially faster than Triplet Margin Loss in preliminary comparisons.Source: D. Training · p. 4
  • Validation loss controls learning-rate reduction, early stopping, and selection of the final model for held-out test evaluation.The learning rate is reduced after 10 epochs without at least 1% validation-loss improvement, and training halts after 20 such epochs.Source: D. Training · p. 4

E. Player Identification

  • Nearest-centroid, nearest-medoid, and k-NN classify embeddings using cosine similarity; the centroid prototype is the normalized mean of a player’s reference embeddings.The medoid instead selects the reference embedding with the highest average similarity to same-player references, while k-NN uses a majority vote.Source: E. Player Identification · p. 4
  • Averaging M query embeddings produces one prediction from sequences spanning 20 seconds, one minute, or two minutes for M = 1, 3, or 6.Reference and query sets use a balanced 50/50 split, repeated across five random seeds without play-session leakage; unknown-player evaluation tests transfer to unseen identities.Source: E. Player Identification · p. 4

A. Experimental Setup

The experiments vary the number of known training players and compare similarity-based classifiers and query-averaging levels. Nearest-centroid classification performs best, while k-NN has substantially higher runtime.

A. Experimental Setup

  • The training-player sweep uses n from 100 to 1000, keeps unknown-player sets fixed, and repeats each setting with five random seeds.Evaluation is macro-averaged because class distribution is imbalanced, with balanced accuracy emphasized for per-player comparison.Source: A. Experimental Setup · pp. 4–5

B. Classification Method Comparison

  • Across k values, k-NN performance changes only marginally; increasing k slightly hurts known-player performance but modestly helps unknown players.The authors attribute the unknown-player trend to noisier embeddings that benefit from majority voting.Source: B. Classification Method Comparison · p. 5
  • Centroid-based classification consistently outperforms medoid and k-NN for known and unknown players at every tested averaging level.Table I covers balanced accuracy across evaluation configurations and averaging levels for n = 1000.Source: B. Classification Method Comparison · p. 5, Table I · p. 5
  • Both centroid and medoid classification finish in under two seconds, while k-NN takes 10.4 to 115.3 seconds depending on M and reference sets.k-NN runtime scales linearly with the number of reference samples.Source: B. Classification Method Comparison · p. 5

C. Scaling of Player Identification

Nearest-centroid identification remains strong for known players as the training set grows, while performance on unseen players improves substantially with more training identities. The results suggest a specialization–generalization tradeoff.

C. Scaling of Player Identification

  • Balanced accuracy and macro F1 closely match across configurations, and low standard deviations indicate stable training and evaluation across data splits.Table II reports both metrics across training-player counts, evaluation configurations, and averaging levels.Source: C. Scaling of Player Identification · p. 5
  • At n = 1000, nearest-centroid balanced accuracy reaches 0.9503 for known players with M = 6 and 0.8786 for unknown players with M = 6.For M = 1, the corresponding values are 0.7899 and 0.6210.Source: C. Scaling of Player Identification · pp. 5–6
  • Unknown-player balanced accuracy rises from 0.2217 at n = 100, M = 1 to 0.6210 at n = 1000, M = 1, and 0.8786 at n = 1000, M = 6.The authors report that useful unseen-player identification emerges as the number of known training players increases.Source: C. Scaling of Player Identification · p. 6
  • Known-player performance declines as n grows, while increasing n brings better unseen-player generalization, suggesting a tradeoff between specialization and generalization.The authors interpret the pattern as learning more transferable aspects of player behavior as the number of known players grows.Source: C. Scaling of Player Identification · p. 6

D. Traditional Baseline

The traditional baseline combines hand-crafted statistical features with conventional classifiers and nearest-centroid classification. PlayerEmbed performs better overall, particularly as the number of known players increases.

D. Traditional Baseline

  • The traditional baseline uses 151-dimensional hand-crafted features, including state frequencies, transitions, durations, value statistics, temporal changes, histograms, and autocorrelations.These features are evaluated with MLP, SVM, XGBoost, and nearest-centroid classification.Source: D. Traditional Baseline · p. 6
  • At n = 1000, the best traditional method reaches balanced accuracy of only 0.3946 with M = 6; PlayerEmbed maintains stronger performance across settings.For n = 100, SVM reaches 0.7951 with M = 6, slightly outperforming MLP; all traditional classifiers benefit from multi-sample averaging.Source: D. Traditional Baseline · p. 6
  • PlayerEmbed substantially outperforms hand-crafted representations, especially in the large-player setting, where traditional-method performance drops substantially.Table III compares balanced accuracy for traditional baselines and PlayerEmbed.Source: D. Traditional Baseline · p. 6

E. Encoder Architecture Evaluation

The evaluation compares CNN-LSTM with Transformer and MLP encoders and tests the effects of removing CNN feature extraction or the projection head. The full CNN-LSTM performs best overall for known and unknown players while training substantially faster than Transformer variants.

E. Encoder Architecture Evaluation

  • Disabling both CNN feature extraction and the projection head decreases balanced accuracy by 30.9% for known and 35.1% for unknown players at M = 1, and by 10.4% and 12.9% at M = 6.Removing either component also substantially reduces performance, especially at M = 1.Source: E. Encoder Architecture Evaluation · pp. 6–7
  • With L = 3 and M = 6, Transformer reaches 0.931 balanced accuracy on known players, while its best unknown-player result is 0.803 with L = 2.Transformer benefits more from multi-sample averaging and trains more slowly; the MLP performs far worse than both Transformer and CNN-LSTM.Source: E. Encoder Architecture Evaluation · pp. 6–7
  • The full CNN-LSTM achieves the best performance for known and unknown players while training substantially faster than Transformer variants.CNN-LSTM averages 4.5 hours of training, compared with 8.1, 22.1, and 28.6 hours for Transformer models with 1, 2, and 3 layers, respectively.Source: E. Encoder Architecture Evaluation · p. 7

F. Embedding Space Analysis

The embedding-space analysis finds a largely smooth shared structure for known and unknown players, alongside separated clusters associated with client modifications and cheat clients. The conclusion reports generalization to unseen identities and identifies broader, more diverse cross-game datasets as future work.

F. Embedding Space Analysis

  • Most player centroids occupy one large, contiguous cluster, with known and unknown players in overlapping regions of an overall smooth, mostly uniform embedding space.The analysis applies UMAP to player centroids and HDBSCAN to identify distinct clusters.Source: F. Embedding Space Analysis · p. 7
  • Four smaller, dense clusters contain 20 to 30 players each and correspond to client modifications, including two clusters associated with cheat clients.Manual gameplay inspection linked two clusters to non-default configurations and two to unusual input behavior, including aim assistance.Source: F. Embedding Space Analysis · p. 7

V. CONCLUSION AND FURTHER WORK

  • The authors conclude that the embedding structure reflects meaningful behavioral patterns beyond player identity and may be useful for other downstream tasks.They report robust identification for known and unknown players.Source: §V · p. 7, F. Embedding Space Analysis · p. 7
  • The reported evaluation covers 1,000 known and 1,000 unknown players; scaling beyond this and collecting more diverse datasets across games remain future work.The authors expect transfer to other games where raw input sequences are available.Source: §V · p. 7

Related papers

Loading 2610.32727v1…