Yan kanal saldırılarının derin öğrenme tabanlı analizi
Deep learning-based analysis of sidechannel attacks
- Tez No: 1015579
- Danışmanlar: PROF. DR. ÜNAL ÇAVUŞOĞLU
- Tez Türü: Yüksek Lisans
- Konular: Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol, Computer Engineering and Computer Science and Control
- Anahtar Kelimeler: Eliptik Eğri Kriptografisi, Yan Kanal Saldırıları, Güç Analizi, Derin Öğrenme, GPAM, Gömülü Sistem Güvenliği, Elliptic Curve Cryptography, Side-Channel Attacks, Power Analysis, Deep Learning, Gaussian Process Assisted Model, Embedded System Security
- Yıl: 2025
- Dil: Türkçe
- Üniversite: Sakarya Üniversitesi
- Enstitü: Fen Bilimleri Enstitüsü
- Ana Bilim Dalı: Bilgisayar Bilimleri Ana Bilim Dalı
- Bilim Dalı: Siber Güvenlik Bilim Dalı
- Sayfa Sayısı: Belirtilmemiş.
Özet
Günümüzde kriptografik algoritmalar, dijital sistemlerin güvenliğini sağlamak amacıyla yaygın biçimde kullanılmaktadır. Açık anahtarlı kriptografi alanında önemli bir yere sahip olan Eliptik Eğri Kriptografisi (Eliptik Eğri Kriptografisi – ECC), sunduğu yüksek güvenlik seviyesi ve düşük anahtar uzunluğu gereksinimi sayesinde özellikle gömülü sistemler, IoT cihazları ve kaynak kısıtlı donanım platformlarında tercih edilmektedir. Ancak kriptografik algoritmaların pratikte fiziksel donanımlar üzerinde çalıştırılması, bu sistemleri matematiksel güvenliğin ötesinde yeni tehditlere açık hâle getirmektedir. Bu tehditlerin başında, kriptografik işlemler sırasında ortaya çıkan fiziksel sızıntılardan yararlanan yan kanal saldırıları gelmektedir. Yan kanal saldırıları, bir kriptografik algoritmanın gizli anahtarını doğrudan matematiksel olarak çözmek yerine, sistemin güç tüketimi, zamanlama bilgisi veya elektromanyetik yayınım gibi yan etkilerini analiz etmeye dayanmaktadır. Bu saldırı türleri arasında güç analizi saldırıları, uygulanabilirliği ve pratikteki başarısı nedeniyle öne çıkmaktadır. Özellikle ECC algoritmalarında yer alan skaler çarpma işlemleri, gizli anahtar ile doğrudan ilişkili olması sebebiyle güç analizi saldırıları açısından kritik bir hedef oluşturmaktadır. Son yıllarda klasik istatistiksel güç analizi yöntemlerinin yanı sıra, makine öğrenmesi ve derin öğrenme tabanlı yaklaşımlar yan kanal analizinde yaygın olarak kullanılmaya başlanmıştır. Derin öğrenme modelleri, gürültülü ve yüksek boyutlu güç tüketimi verileri içerisindeki karmaşık örüntüleri otomatik olarak öğrenebilme yetenekleri sayesinde geleneksel yöntemlere kıyasla daha başarılı sonuçlar sunmaktadır. Bu bağlamda, olasılıksal modelleme kabiliyeti ile öne çıkan Gaussian Process Assisted Model (GPAM) gibi hibrit yaklaşımlar, yan kanal analizinde dikkat çekici bir alternatif oluşturmaktadır. Bu tez çalışmasında, STM32F serisi bir mikrodenetleyici üzerinde çalışan ECC uygulamalarından elde edilen gerçek güç tüketimi verileri kullanılarak, ECC algoritmalarının yan kanal saldırılarına karşı dayanıklılığı incelenmiştir. Güç ölçümleri, ChipWhisperer Lite donanımı kullanılarak gerçekleştirilmiş ve elde edilen güç izleri yalnızca güç analizi bağlamında değerlendirilmiştir. Toplanan veriler, derin öğrenme tabanlı GPAM modeli ile analiz edilerek, güç tüketimi ile gizli anahtar arasındaki ilişki ortaya konulmaya çalışılmıştır. Çalışma kapsamında simülasyon verileri yerine fiziksel ölçümlere dayalı gerçek donanım verilerinin kullanılması, elde edilen sonuçların pratik sistemler açısından anlamlı olmasını sağlamaktadır. Elde edilen bulgular, ECC tabanlı gömülü sistemlerin yalnızca matematiksel güvenlik açısından değil, uygulama ve donanım düzeyinde de değerlendirilmesi gerektiğini göstermektedir. Bu yönüyle tez, yapay zekâ destekli yan kanal analizi alanında ECC özelinde deneysel ve özgün bir katkı sunmayı hedeflemektedir.
Özet (Çeviri)
This thesis presents an in-depth and systematic investigation into the practical security of Elliptic Curve Cryptography (ECC) implementations when exposed to side-channel attacks (SCAs), with a particular emphasis on power analysis techniques and the growing impact of deep learning-based methodologies. While ECC is widely acknowledged as a secure and efficient public-key cryptographic scheme primarily due to the computational hardness of the elliptic curve discrete logarithm problem (ECDLP) its real-world deployment on physical devices introduces a range of vulnerabilities that are not captured by purely theoretical security models. These vulnerabilities stem from unintended physical leakages produced during computation, including fluctuations in power consumption, timing discrepancies, electromagnetic emanations, and even acoustic signals. Such leakages provide adversaries with alternative pathways to infer secret information without directly attacking the underlying mathematical problem. Unlike classical cryptanalysis, which focuses on exploiting weaknesses in the mathematical structure of cryptographic algorithms, side-channel attacks operate by observing the physical behavior of a device as it executes cryptographic operations. This distinction makes SCAs particularly powerful and dangerous in practice, as they effectively bypass the theoretical guarantees that underpin modern cryptographic systems. Among the various types of side-channel attacks, power analysis has emerged as one of the most practical, accessible, and extensively studied techniques. Simple Power Analysis (SPA) involves visually inspecting power traces to identify patterns corresponding to specific operations, enabling attackers to infer secret-dependent control flow. More advanced techniques, such as Differential Power Analysis (DPA) and Correlation Power Analysis (CPA), utilize statistical methods applied across large numbers of traces to extract subtle leakage correlated with secret data. These approaches are especially effective in embedded systems and Internet of Things (IoT) devices, where attackers may have physical access and can perform repeated measurements under controlled conditions. ECC implementations are particularly vulnerable to side-channel attacks due to the nature of their core operation: scalar multiplication. This operation computes a point multiplication 𝑘⋅𝑃, where 𝑘 is a secret scalar and 𝑃 is a public point on the elliptic curve. Scalar multiplication is typically implemented as a sequence of point doubling and point addition operations, often executed in a manner that depends on the binary representation of the secret key. As a result, variations in execution flow and intermediate values lead to observable differences in power consumption, which can be exploited to recover key bits. Even implementations that attempt to standardize execution patterns may still exhibit subtle leakage due to microarchitectural effects, data-dependent memory accesses, or imperfect countermeasures. To mitigate these vulnerabilities, a variety of countermeasures have been proposed, including masking techniques that randomize intermediate values, algorithmic randomization such as scalar blinding, and constant-time implementations that aim to eliminate data-dependent control flow. While these approaches improve resistance against classical attacks, they often incur performance overhead and may not fully eliminate leakage, particularly against more sophisticated adversaries equipped with advanced analysis techniques. In recent years, the field of side-channel analysis has undergone a significant transformation with the integration of machine learning and deep learning approaches. Traditional methods like DPA and CPA rely on predefined leakage models such as Hamming weight or Hamming distance that assume a simplified linear relationship between processed data and observed leakage. However, real-world leakage is often highly non-linear and influenced by noise, hardware variability, and the presence of countermeasures. Deep learning methods address these limitations by learning complex relationships directly from raw measurement data, without requiring explicit modeling of the leakage process. Convolutional Neural Networks (CNNs) have been widely adopted in this context due to their ability to automatically extract local features and tolerate noise in power traces. They have demonstrated strong performance in various side-channel scenarios, including masked implementations. However, CNNs are inherently limited in capturing long-range dependencies across time-series data, which can be critical when analyzing long cryptographic operations such as ECC scalar multiplication. To overcome these limitations, recent research has explored the use of Transformer-based architectures, which leverage self-attention mechanisms to model global dependencies within input sequences. Transformers can dynamically focus on relevant parts of a power trace, regardless of their position, making them particularly well-suited for high-dimensional and temporally complex side-channel data. One prominent example is the Generalized Power Analysis Model (GPAM), which utilizes Transformer encoder blocks to process raw power traces without requiring preprocessing steps such as alignment, filtering, or manual feature extraction. GPAM represents a significant advancement by offering improved scalability, flexibility, and generalization across different devices and cryptographic implementations. This thesis evaluates the effectiveness of deep learning-based side-channel attacks using the GPAM framework on an ECC implementation deployed on an STM32F303 microcontroller. Unlike many prior studies that rely on publicly available datasets or controlled benchmarking environments, this work adopts a custom experimental setup designed to reflect realistic attack conditions. Power traces are collected using the ChipWhisperer-Lite platform, which measures the device's power consumption through a shunt resistor and digitizes the signal using a high-resolution analog-to-digital converter. A hardware trigger integrated into the firmware ensures precise alignment by capturing only the relevant portion of execution, specifically the scalar multiplication phase of the ECDSA algorithm. The resulting dataset consists of tens of thousands of high-resolution power traces, each containing 43,350 sampling points and corresponding to a unique secret key. To ensure a rigorous and unbiased evaluation, no key is reused across training, validation, and test sets, thereby preventing data leakage and ensuring that the model must generalize to unseen keys. The dataset is stored in TFRecord format, enabling efficient data handling and seamless integration with modern deep learning pipelines. A distinctive contribution of this work is the deliberate use of raw, unprocessed power traces. While many previous studies rely on preprocessing techniques such as noise filtering, trace alignment, or dimensionality reduction, this thesis evaluates the ability of Transformer-based models to learn directly from noisy and unstructured real-world data. This approach better reflects realistic attack scenarios, where preprocessing may be limited, unreliable, or impractical. The experimental evaluation explores multiple configurations, including varying trace lengths and different target representations of the secret key, such as byte-level and bit-level predictions. Initial experiments using shorter traces and byte-level targets yield limited success, indicating that the available leakage is insufficient for reconstructing higher-level key representations. However, when the prediction task is simplified to individual bits, the model demonstrates the ability to successfully learn certain key bits, particularly the least significant bit. This finding suggests that leakage is not uniformly distributed across the key and that certain bits are inherently more exposed due to implementation-specific characteristics. Further experiments involving longer traces introduce additional challenges, most notably overfitting. Although the model achieves high accuracy on the training dataset, its performance on validation and unseen test data remains close to random guessing. This indicates that the model is memorizing patterns specific to the training traces rather than learning generalizable leakage features. The high dimensionality of the traces, combined with noise and limited dataset diversity, makes it difficult for the model to extract robust and transferable information. These observations highlight fundamental trade-offs in deep learning-based side-channel analysis. Increasing the trace length provides more information but also amplifies noise and computational complexity. Similarly, predicting more complex targets, such as full bytes, requires stronger and more consistent leakage signals. The interplay between dataset size, trace quality, model architecture, and target representation plays a critical role in determining the success of an attack. Overall, this thesis makes several key contributions. It demonstrates the feasibility of applying Transformer-based architectures to real-world ECC implementations, provides empirical insights into the challenges of generalization and overfitting in high-dimensional side-channel data, and reinforces the importance of evaluating cryptographic security at the implementation level rather than relying solely on theoretical guarantees. Furthermore, it highlights the need for robust and carefully designed countermeasures, including masking, constant-time execution, noise injection, and hardware-level protections, particularly in resource-constrained environments such as embedded systems and IoT devices. In addition, the study underscores the importance of developing standardized benchmarks and diverse datasets to improve the reproducibility and comparability of deep learning-based side-channel attacks. As machine learning techniques continue to evolve, their integration into side-channel analysis will likely lead to increasingly powerful and automated attack strategies, necessitating a corresponding advancement in defensive techniques. In conclusion, this work bridges the gap between theoretical cryptography and practical security by demonstrating how physical leakages in ECC implementations can be effectively exploited using advanced deep learning techniques. It contributes to a deeper understanding of real-world vulnerabilities, exposes the limitations of current defenses, and provides a strong foundation for future research on secure cryptographic implementations, advanced attack methodologies, and resilient system design.
Benzer Tezler
- Analysis of spectre attacks and development of AI-based detection mechanism
Spectre saldırılarının analizi ve YZ-tabanlı tespit mekanizması geliştirilmesi
HATİCE AKTAŞ AYDIN
Doktora
İngilizce
2025
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolSivas Bilim ve Teknoloji ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
DR. ÖĞR. ÜYESİ GÜLAY YALÇIN ALKAN
- Cumhuriyet Üniversitesi Hastanesindeki lavman uygulamaları ile hastaların bu uygulamalara ilişkin duygu ve düşünceleri
Başlık çevirisi yok
SULTAN ALAN ÇİÇEK
Yüksek Lisans
Türkçe
1987
HemşirelikCumhuriyet ÜniversitesiHemşirelik Ana Bilim Dalı
YRD. DOÇ. DR. MELİHA ATALAY
- Bir pamuk iplikhanesinde fiili çalışma metodlarının incelenmesi ve mevcut işlemlerin standart sürelerinin saptanması
Başlık çevirisi yok
MÜRŞİDE KESEROĞLU
Yüksek Lisans
Türkçe
1985
İşletmeEge ÜniversitesiTekstil Mühendisliği Ana Bilim Dalı
DOÇ. DR. MUSTAFA NAZMİ ERCAN
- Birleşik ve yan ürünlerde maliyet hacim-kar analizleri: bir imalat işletmesinde uygulama denemesi
Başlık çevirisi yok
GÜLİN GÜRDAL