SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.07567 · 2026

Revisiting the LiRA Membership Inference Attack Under Realistic Assumptions

Josep Domingo-Ferrer, David Sánchez, Najeeb Jebreel, Mona Khalil

arXiv · PDF · Open in the Atlas

Code that ran

19 of 23 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 4 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 15 failures of our instrument; 4 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2603.07567.json · badge: /api/ran/2603.07567.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
najeebjebreel/lira_analysis canonical (classifier label)
classifier basis: census
19 ran · 4 no recorded run
FunctionStatusWhere it lives
collect_topk_images Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/compose_top_vulnerable.py
code served (permissive licence) · get_code("30c1c468615ed8aa")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import shutil
from pathlib import Path

# Create a temporary directory structure that mimics the expected layout
temp_dir = tempfile.mkdtemp()
repo_root = Path(temp_dir) / "repo"
dest = Path(temp_dir) / "dest"

repo_root.mkdir(parents=True, exist_ok=True)

# Create the analysis_results structure
analysis_results = repo_root / 'analysis_results'
analysis_results.mkdir(parents=True, exist_ok=True)

# Create figures subdirectory
figures_root = analysis_results / 'figures'
figures_root.mkdir(parents=True, exist_ok=True)

# Create some subdirectories in analysis_results (not under figures)
subdir1 = analysis_results / 'subdir1'
subdir1.mkdir(parents=True, exist_ok=True)

# Create dummy PNG files that match the pattern 'top*vulnerable*.png'
# These should be OUTSIDE of figures_root to be collected
(subdir1 / 'top_vulnerable_001.png').touch()
(subdir1 / 'top_vulnerable_002.png').touch()

# Create another subdir with matching files
subdir2 = analysis_results / 'subdir2'
subdir2.mkdir(parents=True, exist_ok=True)
(subdir2 / 'top_results_vulnerable_003.png').touch()

# Create files in figures_root that should be skipped (relative_to succeeds)
(figures_root / 'top_vulnerable_skip.png').touch()

ARGS = {
    'repo_root': repo_root,
    'dest': dest,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_precision Ran
ran, honoured the contract we drafted
najeebjebreel/lira_analysis/comprehensive_analysis/metrics.py
code served (permissive licence) · get_code("54526585121a623a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Construct arguments for compute_precision
# The function takes three parameters: tpr, fpr, and prior (with default 0.5)

ARGS = {
    "tpr": 0.9,           # True positive rate: 90% of members correctly identified
    "fpr": 0.01,          # False positive rate: 1% of non-members incorrectly flagged
    "prior": 0.5,         # Prior probability of membership: 50% (balanced dataset)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_roc_metrics Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/metrics.py
code served (permissive licence) · get_code("99a772f122b776db")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Set random seed for reproducibility
np.random.seed(0)

# Create sample attack scores (continuous values, higher = more likely member)
# Using 10 samples for a small but meaningful dataset
scores = np.array([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.85, 0.95], dtype=np.float32)

# Create ground truth membership labels (0 = non-member, 1 = member)
# Mix of both classes to avoid degenerate case
labels = np.array([0, 0, 0, 0, 0, 1, 1, 1, 1, 1], dtype=np.int32)

ARGS = {
    "scores": scores,
    "labels": labels,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_tpr_at_fpr Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/metrics.py
code served (permissive licence) · get_code("cf5ec905e15dd612")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create sample ROC curve data
# fpr: False positive rates (should be monotonically increasing)
fpr = np.array([0.0, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0])

# tpr: True positive rates (should be monotonically increasing)
tpr = np.array([0.0, 0.3, 0.5, 0.65, 0.8, 0.85, 0.95, 1.0])

# thresholds: Decision thresholds (should be monotonically decreasing)
thresholds = np.array([np.inf, 0.9, 0.7, 0.5, 0.3, 0.2, 0.1, 0.0])

# target_fpr: We want to find TPR at FPR <= 0.01
target_fpr = 0.01

ARGS = {
    "fpr": fpr,
    "tpr": tpr,
    "thresholds": thresholds,
    "target_fpr": target_fpr,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
create_output_directory Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/run_analysis.py
code served (permissive licence) · get_code("ca73aedeabad1c37")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned PosixPath. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from pathlib import Path
import tempfile

# Create a temporary directory to serve as base_dir
base_dir = Path(tempfile.gettempdir()) / "test_output_base"

# Create an exp_path with at least 4 parts to trigger the dataset/model/run_name extraction
# The path structure: .../dataset_name/model_name/run_name
exp_path = Path("/some/experiment/my_dataset/my_model/run_001")

ARGS = {
    "base_dir": base_dir,
    "exp_path": exp_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
fpr_to_tag Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/reproducibility_analysis.py
code served (permissive licence) · get_code("e214cc0b45e0daee")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

def fpr_to_tag(fpr: float) -> str:
    pct = fpr * 100.0
    value = f'{pct:.3f}'.rstrip('0').rstrip('.')
    return value.replace('.', 'p') + 'pct'

ARGS = {
    'fpr': 0.05
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
fpr_to_tag Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/run_analysis.py
code served (permissive licence) · get_code("408beb717e5ac9fd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# Construct valid arguments for fpr_to_tag
# The function takes a single parameter: fpr (float)
# It converts the float to a percentage, replaces '.' with 'p', and appends 'pct'

ARGS = {
    'fpr': 0.05
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
freedman_diaconis_bins Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/plot_benchmark_distribution.py
code served (permissive licence) · get_code("7f54c1a69ee2722e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create a simple numpy array with varied values for testing
# The function expects a numpy array and will compute bins based on the Freedman-Diaconis rule
values = np.array([1.2, 2.5, 3.1, 4.7, 5.2, 6.8, 7.1, 8.3, 9.5, 10.2, 
                    11.1, 12.4, 13.7, 14.2, 15.8, 16.3, 17.9, 18.5, 19.1, 20.0])

ARGS = {
    "values": values,
    "max_bins": 80,
    "min_bins": 50
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_attack_methods Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/attack.py
code served (permissive licence) · get_code("970e527fa55fb4bf")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# The function takes a config dict and extracts attack methods from it
# It accesses config.get('attack', {}).get('method', [])
# and checks if the result is a string or converts it to a list

# Test case 1: config with attack.method as a string
config_string = {
    'attack': {
        'method': 'fgsm'
    }
}

# Test case 2: config with attack.method as a list
config_list = {
    'attack': {
        'method': ['fgsm', 'pgd', 'cw']
    }
}

# Test case 3: config with missing attack key (should return empty list)
config_empty = {}

# Test case 4: config with attack but no method key
config_no_method = {
    'attack': {}
}

# Use the first config as the primary test case
ARGS = {
    'config': config_string
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_experiment_info Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/analysis_utils.py
code served (permissive licence) · get_code("33f9972a2979215a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import yaml
from pathlib import Path

# Create a temporary directory structure for the experiment
temp_dir = tempfile.TemporaryDirectory()
temp_path = Path(temp_dir.name)

# Create nested directory structure matching expected format
# Format: experiments/dataset/model/config_name
experiment_path = temp_path / "experiments" / "cifar10" / "resnet18" / "experiment_001"
experiment_path.mkdir(parents=True, exist_ok=True)

# Create train_config.yaml
train_config = {
    'training': {
        'num_shadow_models': 5,
        'epochs': 100,
    },
    'model': {
        'architecture': 'resnet18',
    },
    'dataset': {
        'name': 'cifar10',
    }
}

train_config_path = experiment_path / 'train_config.yaml'
with open(train_config_path, 'w') as f:
    yaml.dump(train_config, f)

# Create attack_config.yaml
attack_config = {
    'attack': {
        'type': 'membership_inference',
        'threshold': 0.5,
    }
}

attack_config_path = experiment_path / 'attack_config.yaml'
with open(attack_config_path, 'w') as f:
    yaml.dump(attack_config, f)

# Define arguments for get_experiment_info
ARGS = {
    'experiment_dir': str(experiment_path)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
yaml — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_experiment_config Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/analysis_utils.py
code served (permissive licence) · get_code("9ee0cfeab3854a74")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import yaml
from pathlib import Path

# Create temporary directory with test config files
temp_dir = tempfile.mkdtemp()
experiment_dir = Path(temp_dir)

# Create train_config.yaml
train_config = {
    'training': {
        'num_shadow_models': 5,
        'epochs': 10,
        'batch_size': 32,
        'learning_rate': 0.001
    },
    'model': {
        'architecture': 'resnet18',
        'pretrained': False
    }
}

train_config_path = experiment_dir / 'train_config.yaml'
with open(train_config_path, 'w') as f:
    yaml.dump(train_config, f)

# Create attack_config.yaml
attack_config = {
    'attack': {
        'type': 'membership_inference',
        'threshold': 0.5,
        'num_queries': 1000
    },
    'defense': {
        'method': 'differential_privacy',
        'epsilon': 1.0
    }
}

attack_config_path = experiment_dir / 'attack_config.yaml'
with open(attack_config_path, 'w') as f:
    yaml.dump(attack_config, f)

ARGS = {
    'experiment_dir': experiment_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
yaml — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_pooled_thresholds Ran
ran, honoured the contract we drafted
najeebjebreel/lira_analysis/comprehensive_analysis/threshold_distribution.py
code served (permissive licence) · get_code("67d36c79acbea5e2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (3,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path
import pandas as pd
import numpy as np

# Create temporary CSV files with threshold data
temp_dir = tempfile.TemporaryDirectory()
temp_path = Path(temp_dir.name)

# Create first CSV file
csv_path_1 = temp_path / "thresholds_1.csv"
df_1 = pd.DataFrame({
    'mode': ['target', 'target', 'baseline'],
    'attack': ['fgsm', 'fgsm', 'pgd'],
    'target_fpr': [0.1, 0.1, 0.1],
    'prior': [0.01, 0.01, 0.01],
    'threshold': [0.5, 0.6, 0.7]
})
df_1.to_csv(csv_path_1, index=False)

# Create second CSV file
csv_path_2 = temp_path / "thresholds_2.csv"
df_2 = pd.DataFrame({
    'mode': ['target', 'target', 'baseline'],
    'attack': ['fgsm', 'pgd', 'fgsm'],
    'target_fpr': [0.1, 0.2, 0.1],
    'prior': [0.01, 0.01, 0.02],
    'threshold': [0.55, 0.65, 0.75]
})
df_2.to_csv(csv_path_2, index=False)

ARGS = {
    'csv_paths': [csv_path_1, csv_path_2],
    'attack': 'fgsm',
    'target_fpr': 0.1,
    'prior': 0.01
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (3,) — shape (3,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_thresholds_filtered Ran
ran, honoured the contract we drafted
najeebjebreel/lira_analysis/comprehensive_analysis/threshold_distribution.py
code served (permissive licence) · get_code("a6bc219cba07af29")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 1, dtype family float, shape (2,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import pandas as pd
import numpy as np
from pathlib import Path
import tempfile
import os

# Create a temporary CSV file with test data
temp_dir = tempfile.mkdtemp()
csv_path = Path(temp_dir) / "thresholds.csv"

# Create a DataFrame with the required columns
df = pd.DataFrame({
    'mode': ['target', 'target', 'other', 'target'],
    'attack': ['membership', 'membership', 'membership', 'other_attack'],
    'target_fpr': [0.01, 0.01, 0.01, 0.01],
    'prior': [0.01, 0.01, 0.01, 0.01],
    'threshold': [0.5, 0.6, 0.7, 0.8]
})

df.to_csv(csv_path, index=False)

ARGS = {
    'csv_path': csv_path,
    'attack': 'membership',
    'target_fpr': 0.01,
    'prior_equals': 0.01,
    'mode_equals': 'target'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (2,) — shape (2,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
mean_std Ran
ran, honoured the contract we drafted
najeebjebreel/lira_analysis/comprehensive_analysis/rank_stability_utils.py
code served (permissive licence) · get_code("2aa1b93743d08c9d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Construct a valid numpy array with a mix of finite and non-finite values
arr = np.array([1.0, 2.5, 3.7, np.nan, 5.2, np.inf, 7.1, -np.inf, 9.0])

ARGS = {
    "arr": arr
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_columns Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/loss_ratio_tpr.py
code served (permissive licence) · get_code("b4263be4e1529343")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned DataFrame. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import pandas as pd

# Create a sample DataFrame with columns that the function will rename
df = pd.DataFrame({
    'Loss_Ratio': [0.1, 0.2, 0.3],
    'loss_ratio': [0.15, 0.25, 0.35],
    'TPR_0.1FPR': [0.5, 0.6, 0.7],
    'tpr': [0.55, 0.65, 0.75],
    'TPR@0.1%FPR': [0.52, 0.62, 0.72],
    'other_column': [1, 2, 3]
})

ARGS = {
    'df': df
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DataFrame — a DataFrame output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataFrame output the recorded ndim 2 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_panel Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/plot_benchmark_distribution.py
code served (permissive licence) · get_code("08a2335c3fcc3802")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from pathlib import Path

ARGS = {
    "panel_arg": "MyPanel=/tmp/test_path"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
resolve_config_paths Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/plot_benchmark_distribution.py
code served (permissive licence) · get_code("e4629aee855b31c7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from pathlib import Path

# Create temporary directories for testing
config_dir = Path("/tmp/config")
repo_root_dir = Path("/tmp/repo")
cli_experiments_root_dir = Path("/tmp/experiments")

# Create a minimal config dictionary
config = {
    'experiments_root': 'experiments',
    'panels': [
        {
            'title': 'Panel 1',
            'path': 'results/panel1.png'
        },
        {
            'title': 'Panel 2',
            'path': '/absolute/path/panel2.png'
        }
    ],
    'out': 'output/figures/result.pdf'
}

# Define ARGS with all required parameters
ARGS = {
    'config': config,
    'config_path': config_dir / 'config.yaml',
    'repo_root': repo_root_dir,
    'cli_experiments_root': cli_experiments_root_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
robust_stats_1d Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/rank_stability_utils.py
code served (permissive licence) · get_code("1a7471281a81be03")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# The function robust_stats_1d takes a single parameter: x (np.ndarray)
# It converts x to a float array, filters out non-finite values, and computes statistics
# We need to provide a valid numpy array with some numeric data

x_data = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0])

ARGS = {
    'x': x_data
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
safe_spearman Ran
ran; our contract draft was wrong, not the code
najeebjebreel/lira_analysis/comprehensive_analysis/rank_stability_utils.py
code served (permissive licence) · get_code("34e81e2a77283b09")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
from scipy.stats import spearmanr

def safe_spearman(a: np.ndarray, b: np.ndarray) -> float:
    if a.size < 2:
        return np.nan
    return float(spearmanr(a, b).correlation)

# Setup code
np.random.seed(42)

# Create two numpy arrays with at least 2 elements for valid Spearman correlation
a = np.array([1.5, 2.3, 3.1, 4.8, 5.2])
b = np.array([2.1, 3.5, 3.0, 5.1, 6.0])

ARGS = {
    "a": a,
    "b": b,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_identical_over_thresholds Not yet run
unverified (harvested, no recorded run)
najeebjebreel/lira_analysis/comprehensive_analysis/reproducibility_analysis.py
code served (permissive licence) · get_code("a412bade5c864edb")
How it ran

no invocation recorded.

create_output_directory Not yet run
unverified (harvested, no recorded run)
najeebjebreel/lira_analysis/comprehensive_analysis/analysis_utils.py
code served (permissive licence) · get_code("982573cf8895bdfa")
How it ran

no invocation recorded.

generate_reproducibility_panels Not yet run
unverified (harvested, no recorded run)
najeebjebreel/lira_analysis/comprehensive_analysis/reproducibility_analysis.py
code served (permissive licence) · get_code("0128a2f4eb9aceaf")
How it ran

no invocation recorded.

save_latex_summary Not yet run
unverified (harvested, no recorded run)
najeebjebreel/lira_analysis/comprehensive_analysis/run_analysis.py
code served (permissive licence) · get_code("096a5b0d17325f3e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 23 samples lifted from this paper's linked repositories, 23 carry a function identity and 0 do not. Of the 23 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 23 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Membership inference attacks (MIAs) have become the standard tool for evaluating privacy leakage in machine learning (ML). Among them, the Likelihood-Ratio Attack (LiRA) is widely regarded as the state of the art when sufficient shadow models are available. However, prior evaluations have often overstated the effectiveness of LiRA by attacking models overconfident on their training samples, calibrating thresholds on target data, assuming balanced membership priors, and/or overlooking attack reproducibility. We re-evaluate LiRA under a realistic protocol that (i) trains models using anti-overfitting (AOF) (and transfer learning (TL), when applicable) to reduce overconfidence as it would be desirable in production models; (ii) calibrates decision thresholds from shadow models and data rather than (usually unavailable) target data; (iii) measures positive predictive value (PPV, a.k.a. precision) under shadow-based thresholds and skewed-rather than unrealistically balanced-membership priors (𝜋 ≤ 10%); and (iv) quantifies persample membership reproducibility across different seeds and training variations. In this setting, we find that (a) AOF significantly weakens LiRA and TL further reduces the effectiveness of the attack, while improving model accuracy; (b) with shadow-based thresholds and skewed priors, LiRA's PPV often drops from near-perfect to substantially lower levels, especially under AOF/AOF+TL and for 𝜋 ≤ 10%; and (c) LiRA's thresholded "vulnerable" sets at extremely low FPR exhibit poor reproducibility across runs, while likelihood ratio-based rankings are more stable. These results suggest that (i) LiRA, and likely weaker MIAs, are less effective than previously suggested, and their positive inferences can be less reliable under realistic settings; and (ii) for MIAs to serve as meaningful privacy auditing tools, their evaluation must reflect pragmatic training practices, feasible attacker assumptions, and reproducibility considerations. We release our code at: https://github.com/najeebjebreel/lira_analysis.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2603.07567")
get_code_for_paper("2603.07567")
have("2603.07567")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2603.07567.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2603.07567.svg)](https://syntology.ai/paper/2603.07567)

Connect an agent — have() is free.