SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2010.02432 · ICLR · 2021

A Panda? No, It's a Sloth: Slowdown Attacks on Adaptive Multi-Exit Neural Network Inference

Sanghyun Hong, Yiǧitcan Kaya, Tudor Dumitras, Vlad Modoranu

arXiv · PDF · Open in the Atlas

Code that ran

10 of 17 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; 1 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 1 failure of our instrument; 7 no recorded run), measured 2026-11-21; 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/2010.02432.json · badge: /api/ran/2010.02432.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
sanghyun-hong/deepsloth canonical (classifier label)
classifier basis: census
10 ran · 7 no recorded run
FunctionStatusWhere it lives
PGD Ran
ran on a synthesized input
sanghyun-hong/deepsloth/attacks/PGDs.py
code served (permissive licence) · get_code("de06046620195b86")
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 4) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch
import torch.nn as nn
from torch.autograd import Variable

torch.manual_seed(0)

# Create a simple CNN model that takes input and returns predictions
class SimpleModel(nn.Module):
    def __init__(self, cnn_mode=True):
        super(SimpleModel, self).__init__()
        self.cnn_mode = cnn_mode
        if cnn_mode:
            self.conv = nn.Conv2d(3, 8, kernel_size=3, padding=1)
            self.fc = nn.Linear(8 * 4 * 4, 5)
        else:
            self.fc1 = nn.Linear(48, 16)
            self.fc2 = nn.Linear(16, 5)
    
    def forward(self, x):
        if self.cnn_mode:
            x = self.conv(x)
            x = x.view(x.size(0), -1)
            x = self.fc(x)
            return x
        else:
            x = x.view(x.size(0), -1)
            x = self.fc1(x)
            x = self.fc2(x)
            return [x]

# Create model for CNN mode
model_cnn = SimpleModel(cnn_mode=True)

# Create model for non-CNN mode
model_noncnn = SimpleModel(cnn_mode=False)

# Create input tensors: batch_size=2, 3 channels, 4x4 spatial dims
x = torch.rand(2, 3, 4, 4)

# Create labels
y = torch.randint(0, 5, (2,))

# Create predictions (not used in the function but required as param)
preds = None

# Create loss function
loss_fn = nn.CrossEntropyLoss()

# Define ARGS for CNN mode
ARGS = {
    'x': x,
    'preds': preds,
    'loss_fn': loss_fn,
    'y': y,
    'model': model_cnn,
    'eps': 0.3,
    'steps': 2,
    'gamma': 0.1,
    'norm': 'linf',
    'randinit': False,
    'cuda': False,
    'cnn': True,
    'kwargs': {}
}
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 4) — shape (2, 3, 4, 4) 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
PGD_avg Ran
ran on a synthesized input
sanghyun-hong/deepsloth/attacks/PGDs.py
code served (permissive licence) · get_code("96433a42de458de2")
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 Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch
import torch.nn as nn

torch.manual_seed(0)

# Create a simple model that returns a list of predictions
class SimpleModel(nn.Module):
    def __init__(self):
        super(SimpleModel, self).__init__()
        self.fc = nn.Linear(4, 3)
    
    def forward(self, x):
        # Return a list of 2 predictions for PGD_avg to average
        out = self.fc(x)
        return [out, out]

# Create tensors
batch_size = 2
input_dim = 4
num_classes = 3

x = torch.randn(batch_size, input_dim)
y = torch.tensor([0, 1])

# Create model
model = SimpleModel()

# Create loss function
loss_fn = nn.CrossEntropyLoss()

# Create preds (dummy, not used in function body)
preds = None

ARGS = {
    'x': x,
    'preds': preds,
    'loss_fn': loss_fn,
    'y': y,
    'model': model,
    'eps': 0.3,
    'steps': 2,
    'gamma': 0.1,
    'norm': 'linf',
    'randinit': False,
    'cuda': False,
}
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 Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) 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
convert_num_early_exits_at_each_ic_to_cumulative_dis Ran
ran on a synthesized input
sanghyun-hong/deepsloth/delay_attack_cost.py
code served (permissive licence) · get_code("e0235251ee546bc1")
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
import numpy as np

torch.manual_seed(0)

# Analyze the function:
# - ic_exits: a list/sequence that we iterate over with len() and indexing
# - total_samples: an integer used for division
# The function counts cumulative exits at each intermediate classifier,
# normalizes by total_samples, and returns a list of proportions.

ic_exits = [10, 15, 20, 5]
total_samples = 50

ARGS = {
    "ic_exits": ic_exits,
    "total_samples": total_samples
}
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 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
convert_set_of_early_exit_samples_to_cumulative_dist Ran
ran on a synthesized input
sanghyun-hong/deepsloth/delay_attack_cost.py
code served (permissive licence) · get_code("971a65c04cc88912")
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

# Construct ic_exits: a dictionary mapping layer indices to sets of sample indices
# that exited correctly at that layer
ic_exits = {
    0: {0, 1},           # 2 samples exited correctly at layer 0
    1: {0, 2},           # 2 samples exited correctly at layer 1 (0 already counted)
    2: {1, 3},           # 2 samples exited correctly at layer 2 (1 already counted)
}

# Total number of samples in the dataset
total_samples = 4

ARGS = {
    "ic_exits": ic_exits,
    "total_samples": total_samples,
}
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
feature_reduction_formula Ran
ran, honoured the contract we drafted
sanghyun-hong/deepsloth/utils.py
pointer only (licence: MIT); the stored copy get_code serves, from gorakraj/earlyexit_onnx, is pointer only (licence: NONE) · get_code("f79d8861a3c36477")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from gorakraj/earlyexit_onnx, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_loader Ran
ran; our contract draft was wrong, not the code
sanghyun-hong/deepsloth/model_funcs.py
code served (permissive licence) · get_code("2afaaf23c911b3e5")
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. 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 DummyData:
    def __init__(self):
        # Create small dummy dataloaders
        x_train = torch.randn(2, 4)
        y_train = torch.randint(0, 2, (2,))
        self.train_loader = [(x_train, y_train)]
        
        x_aug = torch.randn(2, 4)
        y_aug = torch.randint(0, 2, (2,))
        self.aug_train_loader = [(x_aug, y_aug)]

data = DummyData()
augment = True

ARGS = {
    "data": data,
    "augment": augment
}
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 — a list 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
get_subsets Ran
ran on a synthesized input
sanghyun-hong/deepsloth/utils.py
code served (permissive licence) · get_code("23687b95e2b07471")
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 2, dtype family int; 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 itertools as it
import torch
import numpy as np

torch.manual_seed(0)

ARGS = {
    "input_list": [1, 2, 3, 4, 5],
    "sset_size": 2
}
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 list, ndim 2, dtype family int — 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
load_from_pickle Ran
ran on a synthesized input
sanghyun-hong/deepsloth/utils.py
code served (permissive licence) · get_code("6a4c4e404ba00cd1")
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 pickle
import tempfile
import os

# Create a temporary pickle file with test data
temp_dir = tempfile.mkdtemp()
test_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "dict"}
}
temp_filepath = os.path.join(temp_dir, "test.pkl")
with open(temp_filepath, 'wb') as f:
    pickle.dump(test_data, f)

ARGS = {
    "filename": temp_filepath
}
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 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
profile Ran
ran on a synthesized input
sanghyun-hong/deepsloth/profiler.py
code served (permissive licence) · get_code("dbcc2e5aa63869de")
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 2, 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
import torch.nn as nn

torch.manual_seed(0)

# Build a small model with various layer types
class SmallModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 8, kernel_size=3, padding=1)
        self.bn1 = nn.BatchNorm2d(8)
        self.relu1 = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=2)
        self.linear = nn.Linear(8 * 8 * 8, 10)
    
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu1(x)
        x = self.maxpool(x)
        x = x.view(x.size(0), -1)
        x = self.linear(x)
        return x

model = SmallModel()
input_size = 16
device = 'cpu'

ARGS = {
    'model': model,
    'input_size': input_size,
    'device': device
}
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 2, 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
update_numpy Ran
ran on a synthesized input
sanghyun-hong/deepsloth/scenarios.py
code served (permissive licence) · get_code("eb2987c7e3d2b22b")
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 (4,) 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 torch
import numpy as np

torch.manual_seed(0)

# The function update_numpy takes:
# - acc: an accumulator (can be None or a numpy array)
# - term: a term to accumulate (numpy array)
# - func: a function that takes a tuple (acc, term) and returns updated value

# Create sample numpy arrays
acc_init = None
term_array = np.array([1.0, 2.0, 3.0, 4.0])

# Define a simple accumulation function
def accumulate_func(args):
    acc, term = args
    return acc + term

ARGS = {
    "acc": acc_init,
    "term": term_array,
    "func": accumulate_func
}
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 ndarray, ndim 1, dtype family float, shape (4,) — shape (4,) 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
PGD_max Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/attacks/PGDs.py
code served (permissive licence) · get_code("1d50c333062f5b9a")
How it ran

no invocation recorded.

accuracy Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/datasets.py
pointer only (licence: MIT); the stored copy get_code serves, from gorakraj/earlyexit_onnx, is pointer only (licence: NONE) · get_code("0ae2381aa39b7e94")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from gorakraj/earlyexit_onnx, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

get_class_wise_lists Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/scenarios.py
code served (permissive licence) · get_code("61f6c47663e10368")
How it ran

no invocation recorded.

get_plot_data_and_auc Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/delay_attack_cost.py
code served (permissive licence) · get_code("4d847ffe4199fc69")
How it ran

no invocation recorded.

load_train_loader Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/datasets.py
code served (permissive licence) · get_code("e5ae7732cd5d8200")
How it ran

no invocation recorded.

load_valid_loader Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/datasets.py
code served (permissive licence) · get_code("e19e45cb0c0659ae")
How it ran

no invocation recorded.

scenario_1_split Not yet run
unverified (harvested, no recorded run)
sanghyun-hong/deepsloth/scenarios.py
code served (permissive licence) · get_code("2a83ddc10a6bd194")
How it ran

no invocation recorded.

Also present in other papers' code

Of 17 samples lifted from this paper's linked repositories, 17 carry a function identity and 0 do not. Of the 17 distinct functions with an identity, 5 are also present in other papers whose code we hold (3 with at least one paper dated earlier, 0 the same month, 4 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.

FunctionAlso present inEarliest-datedOthers named
accuracy 2 other papers
1 dated earlier, 1 later
2002.10025 (2020-02)
Triple Wins: Boosting Accuracy, Robustness and Efficiency Together by Enabling Input-Adaptive Inference
136710691 (2022 (year only))
Meta-GF: Training Dynamic-Depth Neural Networks Harmoniously
feature_reduction_formula 2 other papers
1 dated earlier, 1 later
2002.10025 (2020-02)
Triple Wins: Boosting Accuracy, Robustness and Efficiency Together by Enabling Input-Adaptive Inference
136710691 (2022 (year only))
Meta-GF: Training Dynamic-Depth Neural Networks Harmoniously
get_loader 1 other paper
1 dated earlier
2002.10025 (2020-02)
Triple Wins: Boosting Accuracy, Robustness and Efficiency Together by Enabling Input-Adaptive Inference
get_subsets 1 other paper
1 later
136710691 (2022 (year only))
Meta-GF: Training Dynamic-Depth Neural Networks Harmoniously
profile 1 other paper
1 later
136710691 (2022 (year only))
Meta-GF: Training Dynamic-Depth Neural Networks Harmoniously

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

Recent increases in the computational demands of deep neural networks (DNNs), combined with the observation that most input samples require only simple models, have sparked interest in input-adaptive multi-exit architectures, such as MSDNets or Shallow-Deep Networks. These architectures enable faster inferences and could bring DNNs to low-power devices, e.g., in the Internet of Things (IoT). However, it is unknown if the computational savings provided by this approach are robust against adversarial pressure. In particular, an adversary may aim to slowdown adaptive DNNs by increasing their average inference time-a threat analogous to the denial-of-service attacks from the Internet. In this paper, we conduct a systematic evaluation of this threat by experimenting with three generic multi-exit DNNs (based on VGG16, MobileNet, and ResNet56) and a custom multi-exit architecture, on two popular image classification benchmarks (CIFAR-10 and Tiny ImageNet). To this end, we show that adversarial example-crafting techniques can be modified to cause slowdown, and we propose a metric for comparing their impact on different architectures. We show that a slowdown attack reduces the efficacy of multi-exit DNNs by 90-100%, and it amplifies the latency by 1.5-5× in a typical IoT deployment. We also show that it is possible to craft universal, reusable perturbations and that the attack can be effective in realistic black-box scenarios, where the attacker has limited knowledge about the victim. Finally, we show that adversarial training provides limited protection against slowdowns. These results suggest that further research is needed for defending multi-exit architectures against this emerging threat. Our code is available at https://github.com/sanghyun-hong/deepsloth.

For agents

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

get_harvested_code_for_paper("2010.02432")
get_code_for_paper("2010.02432")
have("2010.02432")

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

curl https://syntology.ai/api/ran/2010.02432.json

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

[![Syntology run record](https://syntology.ai/api/ran/2010.02432.svg)](https://syntology.ai/paper/2010.02432)

Connect an agent — have() is free.