I work at the intersection of trustworthy AI and systems security, where I study how AI
systems behave under the real conditions of security-critical applications. My research reveals gaps between
benchmark performance and real-world reliability through measurement-driven analysis of
deployed systems, and uses these insights to design AI that is robust, adaptive, and secure.
I apply this approach to real systems, including malware detectors and customer-service chatbots, to uncover
hidden failure modes and improve AI reliability in practice. My prior work identified the overthinking
pathology in deep networks (3, 7, 11); developed
realistic, security-grounded threat models for machine learning (2, 3, 4, 5, 9, 10, 15);
advanced fairness and privacy techniques (8, 13); and demonstrated how
distribution shift undermines deployed defenses (14, 16). My research
has been featured in
VentureBeat
and
MIT
Technology Review.
Most recently, I led a successful
Amazon Research Award
supporting ongoing work on deployment-aware, adaptable AI for security applications.
I have mentored 20+ undergraduate interns and junior researchers in trustworthy AI (through
NSF REU,
Summer at MC2,
and the ACTION Institute), contributing to multiple top-venue
publications
(7, 9, 13, 16) and six PhD
placements in leading programs.
I go by Can, pronounced like 'John' (d͡ʒan).
News
09/25/2025 - 1 new paper in IEEE S&P'26! On prompt
injection risks in third-party AI chatbot plugins on the web! (16).
01/23/2025 - 1 new paper in NAACL'25! On poisoning attacks
against large language models that get them to produce copyrighted content! (15).
01/22/2025 - 1 new paper in SaTML'25! On demystfying the
challenges and pitfalls in behavior-based malware detection in the real world (14). Read more about this project on our [Project Website].
01/01/2024 - 2 new papers in ICLR'24! On enabling certified
robustness for adversarial attacks against raw-byte-based malware detectors (12) and studying the tensions between group robustness and resilience against
poisoning attacks (13).
Prompt injection attacks pose a critical threat to large language models (LLMs), with prior work focusing
on cutting-edge LLM applications like personal copilots. In contrast, simpler LLM applications, such as
customer service chatbots, are widespread on the web, yet their security posture and exposure to such
attacks remain poorly understood. These applications often rely on third-party chatbot plugins that act as
intermediaries to commercial LLM APIs, offering non-expert website builders intuitive ways to customize
chatbot behaviors. To bridge this gap, we present the first large-scale study of 17 third-party chatbot
plugins used by over 10,000 public websites, uncovering previously unknown prompt injection risks in
practice. First, 8 of these plugins (used by 8,000 websites) fail to enforce the integrity of the
conversation history transmitted in network requests between the website visitor and the chatbot. This
oversight amplifies the impact of direct prompt injection attacks by allowing adversaries to forge
conversation histories (including fake system messages), boosting their ability to elicit unintended
behavior (e.g., code generation) by 3–8x. Second, 15 plugins offer tools, such as web-scraping, to enrich
the chatbot's context with website-specific content. However, these tools do not distinguish the website's
trusted content (e.g., product descriptions) from untrusted, third-party content (e.g., customer reviews),
introducing a risk of indirect prompt injection. Notably, we found that ~13% of e-commerce websites have
already exposed their chatbots to third-party content. We systematically evaluate both vulnerabilities
through controlled experiments grounded in real-world observations, focusing on factors such as system
prompt design and the underlying LLM. Our findings show that many plugins adopt insecure practices that
undermine the built-in LLM safeguards.
As the capabilities of large language models (LLMs) continue to expand, their usage has become
increasingly prevalent. However, as reflected in numerous ongoing lawsuits regarding LLM-generated content,
addressing copyright infringement remains a significant challenge. In this paper, we introduce
PoisonedParrot: the first stealthy data poisoning attack that induces an LLM to generate copyrighted content
even when the model has not been directly trained on the specific copyrighted material. PoisonedParrot
integrates small fragments of copyrighted text into the poison samples using an off-the-shelf LLM. Despite
its simplicity, evaluated in a wide range of experiments, PoisonedParrot is surprisingly effective at
priming the model to generate copyrighted content with no discernible side effects. Moreover, we discover
that existing defenses are largely ineffective against our attack. Finally, we make the first attempt at
mitigating copyright-infringement poisoning attacks by proposing a defense: ParrotTrap. We encourage the
community to explore this emerging threat model further.
Malware detection is a ubiquitous application of Machine Learning (ML) in security. In behavioral malware
analysis, the detector relies on features extracted from program execution traces. The research literature
has focused on detectors trained with features collected from sandbox environments and evaluated on samples
also analyzed in a sandbox. However, in deployment, a malware detector at endpoint hosts often must rely on
traces captured from endpoint hosts, not from a sandbox. Thus, there is a gap between the literature and
real-world needs. We present the first measurement study of the performance of ML-based malware detectors at
real-world endpoints. Leveraging a dataset of sandbox traces and a dataset of in-the-wild program traces, we
evaluate two scenarios: (i) an endpoint detector trained on sandbox traces (convenient and easy to train),
and (ii) an endpoint detector trained on endpoint traces (more challenging to train, since we need to
collect telemetry data). We discover a wide gap between the performance as measured using prior evaluation
methods in the literature—over 90%—vs. expected performance in endpoint detection—about 20% (scenario (i))
to 50% (scenario (ii)). We characterize the ML challenges that arise in this domain and contribute to this
gap, including label noise, distribution shift, and spurious features. Moreover, we show several techniques
that achieve 5–30% relative performance improvements over the baselines. Our evidence suggests that applying
detectors trained on sandbox data to endpoint detection is challenging. The most promising direction is
training detectors directly on endpoint data, which marks a departure from current practice. To promote
progress, we will facilitate researchers to perform realistic detector evaluations against our real-world
dataset.
Group robustness has become a major concern in machine learning (ML) as conventional training paradigms
were found to produce high error on minority groups. Without explicit group annotations, proposed solutions
rely on heuristics that aim to identify and then amplify the minority samples during training. In our work,
we first uncover a critical shortcoming of these methods: an inability to distinguish legitimate minority
samples from poison samples in the training set. By amplifying poison samples as well, group robustness
methods inadvertently boost the success rate of an adversary---e.g., from 0% without amplification to over
97% with it. Notably, we supplement our empirical evidence with an impossibility result proving this
inability of a standard heuristic under some assumptions. Moreover, scrutinizing recent poisoning defenses
both in centralized and federated learning, we observe that they rely on similar heuristics to identify
which samples should be eliminated as poisons. In consequence, minority samples are eliminated along with
poisons, which damages group robustness---e.g., from 55% without the removal of the minority samples to 41%
with it. Finally, as they pursue opposing goals using similar heuristics, our attempt to alleviate the
trade-off by combining group robustness methods and poisoning defenses falls short. By exposing this
tension, we also hope to highlight how benchmark-driven ML scholarship can obscure the trade-offs among
different metrics with potentially detrimental consequences.
Machine Learning (ML) models have been utilized for malware detection for over two decades. Consequently,
this ignited an ongoing arms race between malware authors and antivirus systems, compelling researchers to
propose defenses for malware-detection models against evasion attacks. However, most if not all existing
defenses against evasion attacks suffer from sizable performance degradation and/or can defend against only
specific attacks, which makes them less practical in real-world settings. In this work, we develop a
certified defense, DRSM (De-Randomized Smoothed MalConv), by redesigning the de-randomized smoothing
technique for the domain of malware detection. Specifically, we propose a window ablation scheme to provably
limit the impact of adversarial bytes while maximally preserving local structures of the executables. After
showing how DRSM is theoretically robust against attacks with contiguous adversarial bytes, we verify its
performance and certified robustness experimentally, where we observe only marginal accuracy drops as the
cost of robustness. To our knowledge, we are the first to offer certified robustness in the realm of static
detection of malware executables. More surprisingly, through evaluating DRSM against empirical attacks of
different types, we observe that the proposed defense is empirically robust to some extent against a diverse
set of attacks, some of which even fall out of the scope of its original threat model. In addition, we
collected recent benign raw executables from diverse sources, which will be made public as a dataset called
PACE (Publicly Accessible Collection(s) of Executables) to alleviate the scarcity of publicly available
benign datasets for studying malware detection and provide future research with more representative data of
the time.
Dynamic neural networks (DyNNs) have shown promise for alleviating the high computational costs of
pre-trained language models (PLMs), such as BERT and GPT. Emerging slowdown attacks have shown to inhibit
the ability of DyNNs to omit computation, e.g., by skipping layers that are deemed unnecessary. As a result,
these attacks can cause significant delays in inference speed for DyNNs and may erase their cost savings
altogether. Most research in slowdown attacks has been in the image domain, despite the ever-growing
computational costs---and relevance of DyNNs---in the language domain. Unfortunately, it is still not
understood what language artifacts trigger extra processing in a PLM, or what causes this behavior. We aim
to fill this gap through an empirical exploration of the slowdown effect on language models. Specifically,
we uncover a crucial difference between the slowdown effect in the image and language domains, illuminate
the efficacy of pre-existing and novel techniques for causing slowdown, and report circumstances where
slowdown does not occur. Building on these observations, we propose the first approach for mitigating the
slowdown effect. Our results suggest that slowdown attacks can provide new insights that can inform the
development of more efficient PLMs.
Deep neural networks (DNNs) are known to be highly vulnerable to adversarial examples (AEs) that include
malicious perturbations. Assumptions about the statistical differences between natural and adversarial
inputs are commonplace in many detection techniques. As a best practice, AE detectors are evaluated against
'adaptive' attackers who actively perturb their inputs to avoid detection. Due to the difficulties in
designing adaptive attacks, however, recent work suggests that most detectors have incomplete evaluation. We
aim to fill this gap by designing a generic adaptive attack against detectors: the 'statistical
indistinguishability attack' (SIA). The SIA optimizes a novel objective to craft adversarial examples (AEs)
that follow the same distribution as the natural inputs with respect to DNN representations. Our objective
targets all DNN layers simultaneously as we show that AEs being indistinguishable at one layer might fail to
be so at other layers. The SIA is formulated around evading distributional detectors that inspect a set of
AEs as a whole and is also effective against four individual AE detectors, two dataset shift detectors, and
an out-of-distribution sample detector, curated from published works. This suggests that the SIA can be a
reliable tool for evaluating the security of a range of detectors.
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 quantization-aware 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
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.
Deep learning models often raise privacy concerns as they leak information about their training data.
This leakage enables membership inference attacks (MIA) that can identify whether a data point was in a
model's training set. Research shows that some 'data augmentation' mechanisms may reduce the risk by
combatting a key factor increasing the leakage, overfitting. While many mechanisms exist, their
effectiveness against MIAs and privacy properties have not been studied systematically. Employing two recent
MIAs, we explore the lower bound on the risk in the absence of formal upper bounds. First, we evaluate 7
mechanisms and differential privacy, on three image classification tasks. We find that applying augmentation
to increase the model's utility does not mitigate the risk and protection comes with a utility penalty.
Further, we also investigate why popular label smoothing mechanism consistently amplifies the risk. Finally,
we propose 'loss-rank-correlation' (LRC) metric to assess how similar the effects of different mechanisms
are. This, for example, reveals the similarity of applying high-intensity augmentation against MIAs to
simply reducing the training time. Our findings emphasize the utility-privacy trade-off and provide
practical guidelines on using augmentation to manage the trade-off.
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 slow down 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 sample-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
1.5-5x 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.
Machine learning algorithms are vulnerable to data poisoning attacks. Prior taxonomies that focus on
specific scenarios, e.g., indiscriminate or targeted, have enabled defenses for the corresponding subset of
known attacks. Yet, this introduces an inevitable arms race between adversaries and defenders. In this work,
we study the feasibility of an attack-agnostic defense relying on artifacts that are common to all poisoning
attacks. Specifically, we focus on a common element between all attacks: they modify gradients computed to
train the model. We identify two main artifacts of gradients computed in the presence of poison: (1) their
ell2 norms have significantly higher magnitudes than those of clean gradients, and (2) their orientation
differs from clean gradients. Based on these observations, we propose the prerequisite for a generic
poisoning defense: it must bound gradient magnitudes and minimize differences in orientation. We call this
gradient shaping. As an exemplar tool to evaluate the feasibility of gradient shaping, we use differentially
private stochastic gradient descent (DP-SGD), which clips and perturbs individual gradients during training
to obtain privacy guarantees. We find that DP-SGD, even in configurations that do not result in meaningful
privacy guarantees, increases the model's robustness to indiscriminate attacks. It also mitigates worst-case
targeted attacks and increases the adversary's cost in multi-poison scenarios. The only attack we find
DP-SGD to be ineffective against is a strong, yet unrealistic, indiscriminate attack. Our results suggest
that, while we currently lack a generic poisoning defense, gradient shaping is a promising direction for
future research.
New data processing pipelines and unique network architectures increasingly drive the success of deep
learning. In consequence, the industry considers top-performing architectures as intellectual property and
devotes considerable computational resources to discovering such architectures through neural architecture
search (NAS). This provides an incentive
for adversaries to steal these unique architectures; when used in the cloud, to provide Machine Learning as
a Service (MLaaS), the adversaries also have an opportunity to reconstruct the architectures by exploiting a
range of hardware side-channels. However, it is challenging to reconstruct unique architectures and
pipelines without knowing the computational graph (e.g., the layers, branches or skip connections), the
architectural parameters (e.g., the number of filters in a convolutional layer) or the specific
pre-processing steps (e.g. embeddings). In this paper, we design an algorithm that reconstructs the key
components of a unique deep learning system by exploiting a small amount of information leakage from a cache
side-channel attack, Flush+Reload. We use Flush+Reload to infer the trace of computations and the timing for
each computation. Our algorithm then generates candidate computational graphs from the trace and eliminates
incompatible candidates through a parameter estimation process. We implement our algorithm on PyTorch and
Tensorflow. We demonstrate experimentally that we can reconstruct MalConv, a novel data pre-processing
pipeline for malware detection, and ProxylessNAS-CPU, a novel network architecture for the ImageNet
classification optimized to run
on CPUs, without knowing the architecture family. In both cases, we achieve 0% error. These results suggest
hardware side channels are a practical attack vector against MLaaS, and more efforts should be devoted to
understanding their impact on the security of deep learning systems.
Deep neural networks (DNNs) have been shown to tolerate "brain damage": cumulative changes to the
network's parameters (e.g., pruning, numerical perturbations) typically result in a graceful degradation of
classification accuracy. However, the limits of this natural resilience are not well understood in the
presence of small adversarial changes to the DNN parameters' underlying memory representation, such as
bit-flips that may be induced by hardware fault attacks. We study the effects of bitwise corruptions on 19
DNN models---six architectures on three image classification tasks---and we show that most models have at
least one parameter that, after a specific bit-flip in their bitwise representation, causes an accuracy loss
of over 90%. We employ simple heuristics to efficiently identify the parameters likely to be vulnerable. We
estimate that 40-50% of the parameters in a model might lead to an accuracy drop greater than 10% when
individually subjected to such single-bit perturbations. To demonstrate how an adversary could take
advantage of this vulnerability, we study the impact of an exemplary hardware fault attack, Rowhammer, on
DNNs. Specifically, we show that a Rowhammer enabled attacker co-located in the same physical machine can
inflict significant accuracy drops (up to 99%) even with single bit-flip corruptions and no knowledge of the
model. Our results expose the limits of DNNs' resilience against parameter perturbations induced
by real-world fault attacks. We conclude by discussing possible mitigations and future research directions
towards fault attack-resilient DNNs.
We characterize a prevalent weakness of deep neural networks (DNNs)—overthinking—which occurs when a
network can reach correct predictions before its final layer. Overthinking is computationally wasteful, and
it can also be destructive when, by the final layer, the correct prediction changes into a
misclassification. Understanding overthinking requires studying how each prediction evolves during a
network’s forward pass, which conventionally is opaque. For prediction transparency, we propose the
Shallow-Deep Network (SDN), a generic modification to off-the-shelf DNNs that introduces internal
classifiers. We apply SDN to four modern architectures, trained on three image classification tasks, to
characterize the overthinking problem. We show that SDNs can mitigate the wasteful effect of overthinking
with confidence-based early exits, which reduce the average inference cost by more than 50% and preserve the
accuracy. We also find that the destructive effect occurs for 50% of misclassifications on natural inputs
and that it can be induced, adversarially, with a recent backdooring
attack. To mitigate this effect, we propose a new confusion metric to quantify the internal disagreements
that will likely lead to misclassifications.
Recent results suggest that attacks against supervised machine learning systems are quite effective,
while defenses are easily bypassed by new attacks. However, the specifications for machine learning systems
currently lack precise adversary definitions, and the existing attacks make diverse, potentially unrealistic
assumptions about the strength of the adversary who launches them. We propose the FAIL attacker model, which
describes the adversary's knowledge and control along four dimensions. Our model allows us to consider a
wide range
of weaker adversaries who have limited control and incomplete knowledge of the features, learning algorithms
and training instances utilized. To evaluate the utility of the FAIL model, we consider the problem of
conducting targeted poisoning attacks in a realistic setting: the crafted poison samples must have clean
labels, must be individually and collectively inconspicuous, and must exhibit a generalized form of
transferability, defined by the FAIL model. By taking these constraints into account, we design StingRay, a
targeted poisoning attack that is practical against 4 machine learning applications, which use 3 different
learning algorithms, and can bypass 2 existing defenses. Conversely, we show that a prior evasion attack is
less effective under generalized transferability. Such attack evaluations, under the FAIL adversary model,
may also suggest promising directions for future defenses.
(1) Tudor Dumitras, Yigitcan Kaya, Radu Marginean and Octavian Suciu:
Too Big to FAIL: What You Need to Know Before Attacking a Machine Learning System (International Workshop on Security Protocols'18)
There is an emerging arms race in the field of adversarial machine learning (AML). Recent results suggest
that machine learning (ML) systems are vulnerable to a wide range of attacks; meanwhile, there are no
systematic defenses. In this position paper we argue that to make progress toward such defenses, the
specifications for machine learning systems must include precise adversary definitions—a key requirement in
other fields, such as cryptography or network security. Without common adversary definitions, new AML
attacks risk making
strong and unrealistic assumptions about the adversary’s capabilities. Furthermore, new AML defenses are
evaluated based on their robustness against adversarial samples generated by a
specific attack algorithm, rather than by a general class of adversaries. We propose the FAIL adversary
model, which describes the adversary’s knowledge and control along four dimensions: data Features, learning
Algorithms, training Instances and crafting Leverage. We analyze several common assumptions often implicit,
from the AML literature, and we argue that the FAIL model can represent and generalize the adversaries
considered in these references. The FAIL model allows us to consider a range of adversarial capabilities and
enables systematic comparisons of attacks against ML systems, providing a clearer picture of the security
threats that these attacks raise. By evaluating how much a new AML attack’s success depends on the strength
of the adversary along each of the FAIL dimensions, researchers will be able to reason about the real
effectiveness of the attack. Additionally, such evaluations may suggest
promising directions for investigating defenses against the ML threats.
Computer Engineering Bilkent University
September 2012 - January 2017
Received a comprehensive scholarship that includes a full tuition waiver and a stipend. (2012 - 2017)
Conducted research on ML explanation tools and ML model testing methodologies towards using ML to aid access
and permission management efforts of AWS customers.
