SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2110.13541 · NeurIPS · 2021

Qu-ANTI-zation: Exploiting Quantization Artifacts for Achieving Adversarial Outcomes

Sanghyun Hong, Tudor Dumitras, Michael Panaitescu, Yigitcan Kaya

arXiv · PDF · Open in the Atlas

Code that ran

10 of 12 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (10 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 0 failures of our instrument; 2 no recorded run), measured 2026-11-21; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2110.13541.json · badge: /api/ran/2110.13541.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
secure-ai-systems-group/qu-anti-zation alias (classifier label)
classifier basis: classifier_2025
10 ran · 2 no recorded run
FunctionStatusWhere it lives
AsymmetricQuantizer Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("9a335d5e444ef7f6")
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 an instance of its own class, AsymmetricQuantizer: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

class RangeTracker:
    """Stand-in class that tracks min/max values of inputs"""
    def __init__(self):
        self.min_val = torch.tensor(-2.5)
        self.max_val = torch.tensor(3.7)
    
    def __call__(self, inputs):
        """Called in forward pass to update tracking"""
        pass

bits_precision = 8
range_tracker = RangeTracker()

ARGS = {
    'bits_precision': bits_precision,
    'range_tracker': range_tracker,
}
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 AsymmetricQuantizer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a AsymmetricQuantizer 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
MovingAverageRangeTracker Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("48c45d01e5c1b780")
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 an instance of its own class, MovingAverageRangeTracker: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

# The MovingAverageRangeTracker.__init__ takes: shape, track, momentum
# Looking at the forward method, it uses:
# - self.shape (to determine keep_dims and reduce_dims)
# - self.track (boolean to check if tracking is enabled)
# - self.momentum (used in update_range for exponential moving average)
# - self.min_val and self.max_val (buffers, initialized as None)

# shape should be a tuple/list where some dimensions are 1 (reduce_dims) and others aren't (keep_dims)
# Let's use a shape with mixed dimensions to test the permutation logic
shape = (4, 1, 3, 1)

# track should be a boolean
track = True

# momentum should be a float between 0 and 1
momentum = 0.1

ARGS = {
    'shape': shape,
    'track': track,
    'momentum': momentum,
}
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 MovingAverageRangeTracker — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a MovingAverageRangeTracker 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
QuantizedConv2d Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("b37124bb0e393e91")
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 an instance of its own class, QuantizedConv2d: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'in_channels': 3,
    'out_channels': 16,
    'kernel_size': 3,
    'stride': 1,
    'padding': 0,
    'dilation': 1,
    'groups': 1,
    'bias': True,
}
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 QuantizedConv2d — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a QuantizedConv2d 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
QuantizedLinear Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("d8db79419e17e540")
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 an instance of its own class, QuantizedLinear: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'in_features': 4,
    'out_features': 3,
    'bias': True
}
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 QuantizedLinear — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a QuantizedLinear 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
Quantizer Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("d70c915efcf4d1a4")
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 an instance of its own class, Quantizer: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

# Create a small stand-in RangeTracker class that tracks ranges
class RangeTracker(nn.Module):
    def __init__(self):
        super().__init__()
        self.register_buffer('min_val', torch.tensor(float('inf')))
        self.register_buffer('max_val', torch.tensor(float('-inf')))
    
    def __call__(self, inputs):
        # Track min and max of inputs
        self.min_val = torch.min(self.min_val, inputs.min())
        self.max_val = torch.max(self.max_val, inputs.max())

# Create range tracker instance
range_tracker = RangeTracker()

# Create a minimal Quantizer subclass that implements update_params
class ConcreteQuantizer(nn.Module):
    def __init__(self, bits_precision, range_tracker):
        super().__init__()
        self.bits_precision = bits_precision
        self.range_tracker = range_tracker
        self.register_buffer('scale', torch.tensor(1.0))
        self.register_buffer('zero_point', torch.tensor(0.0))
        self.register_buffer('min_val', torch.tensor(0.0))
        self.register_buffer('max_val', torch.tensor(127.0))
    
    def update_params(self):
        # Simple implementation: set scale and zero_point based on bits
        num_quantized_values = 2 ** self.bits_precision - 1
        self.scale = num_quantized_values / (self.range_tracker.max_val - self.range_tracker.min_val + 1e-8)
        self.zero_point = -self.range_tracker.min_val * self.scale
        self.min_val = 0.0
        self.max_val = float(num_quantized_values)
    
    def quantize(self, inputs):
        outputs = inputs * self.scale - self.zero_point
        return outputs
    
    def round(self, inputs):
        from torch import autograd
        class Round(autograd.Function):
            @staticmethod
            def forward(ctx, inputs):
                return torch.floor(inputs + 0.5)
            
            @staticmethod
            def backward(ctx, grads):
                return grads
        
        outputs = Round.apply(inputs)
        return outputs
    
    def clamp(self, inputs):
        outputs = torch.clamp(inputs, self.min_val, self.max_val)
        return outputs
    
    def dequantize(self, inputs):
        outputs = (inputs + self.zero_point) / self.scale
        return outputs
    
    def forward(self, inputs):
        self.range_tracker(inputs)
        self.update_params()
        outputs = self.quantize(inputs)
        outputs = self.round(outputs)
        outputs = self.clamp(outputs)
        outputs = self.dequantize(outputs)
        return outputs

ARGS = {
    'bits_precision': 8,
    'range_tracker': range_tracker
}
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 Quantizer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Quantizer 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
RangeTracker Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("04d0d5fd405c1f27")
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 an instance of its own class, RangeTracker: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

# Create a minimal concrete subclass of RangeTracker
class ConcreteRangeTracker(nn.Module):
    def __init__(self, shape):
        super().__init__()
        self.shape = shape
        self.register_buffer('min_val', None)
        self.register_buffer('max_val', None)
        self.track = True  # Add track attribute that forward() checks
    
    def update_range(self, min_val, max_val):
        # Concrete implementation
        if self.min_val is None:
            self.min_val = min_val
        else:
            self.min_val = torch.min(self.min_val, min_val)
        
        if self.max_val is None:
            self.max_val = max_val
        else:
            self.max_val = torch.max(self.max_val, max_val)

# Define shape with distinct dimensions
# Using shape (2, 3, 1, 4) - mixed dimensions with one size-1 for testing reduce logic
shape = (2, 3, 1, 4)

ARGS = {
    'shape': shape
}
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 RangeTracker — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a RangeTracker 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
Round Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("9c001f882200270e")
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 an instance of its own class, Round: the run constructed the object and did not call it, so no result was computed. 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
from torch import autograd

torch.manual_seed(0)

class Round(autograd.Function):

    @staticmethod
    def forward(ctx, inputs):
        return torch.floor(inputs + 0.5)

    @staticmethod
    def backward(ctx, grads):
        return grads


# Create a small tensor for testing
inputs = torch.randn(2, 4, requires_grad=True)

ARGS = {
    'inputs': inputs
}
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 Round — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Round 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
SignedQuantizer Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("016074abfad976a6")
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 an instance of its own class, SignedQuantizer: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

class RangeTracker(nn.Module):
    """Stand-in range tracker that tracks min/max of inputs"""
    
    def __init__(self):
        super().__init__()
        self.register_buffer('min_val', torch.tensor(float('inf')))
        self.register_buffer('max_val', torch.tensor(float('-inf')))
    
    def __call__(self, inputs):
        """Update tracked range"""
        self.min_val = torch.min(self.min_val, inputs.min())
        self.max_val = torch.max(self.max_val, inputs.max())

# Create instances for the arguments
bits_precision = 8
range_tracker = RangeTracker()

# Patch the SignedQuantizer to implement update_params
class TestSignedQuantizer(nn.Module):
    def __init__(self, bits_precision, range_tracker):
        super().__init__()
        self.bits_precision = bits_precision
        self.range_tracker = range_tracker
        self.register_buffer('scale', torch.tensor(1.0))
        self.register_buffer('zero_point', torch.tensor(0.0))
        self.register_buffer('min_val', torch.tensor(-(1 << (self.bits_precision - 1))))
        self.register_buffer('max_val', torch.tensor((1 << (self.bits_precision - 1)) - 1))
    
    def update_params(self):
        # Simple implementation: scale based on range
        if self.range_tracker.max_val > self.range_tracker.min_val:
            self.scale = (self.range_tracker.max_val - self.range_tracker.min_val) / (self.max_val - self.min_val)
            self.zero_point = -self.range_tracker.min_val / self.scale
        else:
            self.scale = torch.tensor(1.0)
            self.zero_point = torch.tensor(0.0)
    
    def quantize(self, inputs):
        outputs = inputs * self.scale - self.zero_point
        return outputs
    
    def round(self, inputs):
        return torch.floor(inputs + 0.5)
    
    def clamp(self, inputs):
        outputs = torch.clamp(inputs, self.min_val, self.max_val)
        return outputs
    
    def dequantize(self, inputs):
        outputs = (inputs + self.zero_point) / self.scale
        return outputs
    
    def forward(self, inputs):
        self.range_tracker(inputs)
        self.update_params()
        outputs = self.quantize(inputs)
        outputs = self.round(outputs)
        outputs = self.clamp(outputs)
        outputs = self.dequantize(outputs)
        return outputs

ARGS = {
    'bits_precision': 8,
    'range_tracker': RangeTracker(),
}
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 SignedQuantizer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a SignedQuantizer 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
SymmetricQuantizer Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("8a74c2894fe32bb3")
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 an instance of its own class, SymmetricQuantizer: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

class RangeTracker(nn.Module):
    def __init__(self):
        super().__init__()
        self.register_buffer('min_val', torch.tensor(float('inf')))
        self.register_buffer('max_val', torch.tensor(float('-inf')))
    
    def forward(self, inputs):
        self.min_val = torch.min(self.min_val, torch.min(inputs))
        self.max_val = torch.max(self.max_val, torch.max(inputs))

range_tracker = RangeTracker()
range_tracker.min_val = torch.tensor(-2.5)
range_tracker.max_val = torch.tensor(2.5)

bits_precision = 8

ARGS = {
    'bits_precision': bits_precision,
    'range_tracker': range_tracker
}
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 SymmetricQuantizer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a SymmetricQuantizer 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
UnsignedQuantizer Ran
ran on a synthesized input · constructed an object; no result was computed
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("cc8c55399eb7d028")
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 an instance of its own class, UnsignedQuantizer: the run constructed the object and did not call it, so no result was computed. 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 torch.nn as nn

torch.manual_seed(0)

# RangeTracker is called in forward() and needs to be callable
class RangeTracker:
    def __init__(self):
        self.min_val = float('inf')
        self.max_val = float('-inf')
    
    def __call__(self, inputs):
        # Track the range of inputs
        self.min_val = min(self.min_val, inputs.min().item())
        self.max_val = max(self.max_val, inputs.max().item())

# Create range tracker instance
range_tracker = RangeTracker()

# bits_precision: integer parameter
bits_precision = 8

ARGS = {
    'bits_precision': bits_precision,
    'range_tracker': range_tracker
}
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 UnsignedQuantizer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a UnsignedQuantizer 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
QuantizationEnabler Not yet run
unverified (harvested, no recorded run)
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("19b7026f240ed406")
How it ran

no invocation recorded.

train_w_perturb Not yet run
unverified (harvested, no recorded run)
secure-ai-systems-group/qu-anti-zation/attack_w_lossfn.py
code served (permissive licence) · get_code("8cff3673daceae2c")
How it ran

no invocation recorded.

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 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 12 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 2026-11-21.

Repositories linked to this paper

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

Abstract

Quantization is a popular technique that transforms the parameter representation of a neural network from floating-point numbers into lower-precision ones (e.g., 8-bit integers). It reduces the memory footprint and the computational cost at inference, facilitating the deployment of resource-hungry models. However, the parameter perturbations caused by this transformation result in behavioral disparities between the model before and after quantization. For example, a quantized model can misclassify some test-time samples that are otherwise classified correctly. It is not known whether such differences lead to a new security vulnerability. We hypothesize that an adversary may control this disparity to introduce specific behaviors that activate upon quantization. To study this hypothesis, we weaponize quantizationaware training and propose a new training framework to implement adversarial quantization outcomes. Following this framework, we present three attacks we carry out with quantization: (i) an indiscriminate attack for significant accuracy loss; (ii) a targeted attack against specific samples; and (iii) a backdoor attack for controlling the model with an input trigger. We further show that a single compromised model defeats multiple quantization schemes, including robust quantization techniques. Moreover, in a federated learning scenario, we demonstrate that a set of malicious participants who conspire can inject our quantization-activated backdoor. Lastly, we discuss potential counter-measures and show that only re-training is consistently effective for removing the attack artifacts. Our code is available at

For agents

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

get_harvested_code_for_paper("2110.13541")
get_code_for_paper("2110.13541")
have("2110.13541")

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

curl https://syntology.ai/api/ran/2110.13541.json

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

[![Syntology run record](https://syntology.ai/api/ran/2110.13541.svg)](https://syntology.ai/paper/2110.13541)

Connect an agent — have() is free.