In these last years, Blockchain technologies have been widely used in several application elds to improve data privacy and trustworthiness and security of systems. Although the blockchain is a powerful tool, it is not immune to cyber attacks: for instance, recently (January 2019) a successful 51% attack on Ethereum Classic has revealed security vulnerabilities of its platform. Under a statistical perspective, attacks can be seen as an anomalous observation, with a strong deviation from the regular behavior. Machine Learning is a science whose goal is to learn insights, patterns and outliers within large data repositories; hence, it can be exploit for blockchain attack detection. In this work, we de ne an anomaly detection system based on a encoder-decoder deep learning model, that is trained exploiting aggregate information extracted by monitoring blockchain activities. Experiments on complete historical logs of Ethereum Classic network prove the capability of the our model to e ectively detect the publicly reported attacks. To the best of our knowledge, our approach is the rst one that provides a comprehensive and feasible solution to monitor the security of blockchain transactions.
The Blockchain is largely considered an e ective solution able to ensure security and
trustworthiness. Nevertheless, it can be prone to attacks and security threats, as discussed in [
Roughly, the blockchain is characterized by a global ledger that can record transactions e ciently and permanently through a timed sequence of blocks, called chain. Blocks contain information about the transactions, and each block is added to the chain after a validation process, based on a distributed consensus mechanism. Consensus is reached whenever a sufcient number of nodes validates the block, that is then considered trustworthy. This whole process is logged and it is possible to collect information characterizing the activities within the underlying ledger. It is natural to ask whether such information can be exploited to monitor the process and provide early analysis and detection mechanisms capable of reporting anomalous situations and prospective attacks.
Historically, data analysis techniques have been extensively exploited in the cyber security
domain [
In this paper, we propose an encoder-decoder model that exploits the information collected by the activities within a blockchain, to highlight anomalies in the underlying network activity, that can represent a potential symptom of a forthcoming or current attack on the system integrity. The contribution of the paper is twofold:
We identify a set of properties to be computed from the blockchain logs, which describe the status of the system at every time step.
Next, we devise an unsupervised neural architecture, which takes as input a time series representing the status of the blockchain network for a time window, able to compute a score representing the degree of anomaly exhibited by the time series.
Experiments on the historical logs of the ETC network show that the system is e ective in detecting the reported attacks. To the best of our knowledge, our approach is the rst one that provides a comprehensive and feasible solution to monitor the security of blockchain transactions.
The paper is organized as follows. In section 2 we review the architecture of the ETC network and describe the attacks that exposed its potential vulnerabilities. We also review how machine learning and blockchain technology have been combined in the current literature. Section 3 provides two contributions: rst, it gives details about the structure of the ETC logs and the preprocessing steps applied to engineer the relevant features describing the blockchain status. Next, we discuss about the details of the Encoder/Decoder architecture. Section 4 describes the experiments that demonstrate the e ectiveness of the proposed architecture, and nally in section 5 we discuss open problems and pointers to future directions. 2
The blockchain is a Distributed Ledger Technology [
From a technical perspective, a blockchain is a distributed database shared into a peer-topeer network: data are stored under the form of transactions within speci c structures named blocks. Each block is the collection of all transactions performed in a time window and di erent blocks may contain a di erent number of transactions. Blocks are connected by a cryptographic bond in the chain: a hash function is computed on the concatenation of the signature of the current block with the hash computed on the previous one. By repeating this step for each block the chain is created and the references to the previous block are maintained, as shown in
The security is guaranteed by a consensus mechanism that allows to con rm transactions
and produce new blocks. The most popular consensus algorithm is named "Proof-of-Work"
(PoW), rst adopted in the Bitcoin blockchain [
PoW is considered a robust method, able to reach the Byzantine-fault-tolerant consensus
[
Originally, blockchain was proposed as a technology to support cryptocurrencies. It later evolved with smart contracts, i.e. computer programs performed in decentralized way on Blockchain so that they can be executed and veri ed automatically. ETC implements such a technology, by providing a platform for decentralized applications that run and are regulated on a distributed ledger. Each contract is composed of a sequence of operations whose cost is expressed as gas, i.e., the amount of cryptocurrency that is deemed necessary to execute them. Actually, ETC implements a priority mechanism of the transaction execution. The priority depends on the provided gas, that corresponds to the fee received by the rst miner able to compute the solution for the puzzle associated to a block: the higher the provided gas, the higher the priority.
Blockchain is not immune to attacks. The ETC network exhibits some vulnerabilities [
Signorini et al.[
More similar to our approach is the study proposed in [
In this section we devise a machine learning approach to identify anomalies in the usage of Blockchain-based systems. As discussed before, the approach we propose is essentially unsupervised, due to the lack of labeled data: successful attacks represent very rare events and typically they don't share common patterns. As a consequence, supervised models exhibit poor detection performances.
The proposed anomaly detection system builds a neural encoder-decoder model, capable of summarizing the information about the status of the ledger within a latent space and then to rebuild the original information from this space. The underlying intuition is that, whenever the current status is consistent, the encoding/decoding operation preserves the basic properties of the data. By contrast, anomalous situations exhibit inconsistent values that ultimately result in a failing reconstruction. This happens, for example, when the amount of ether (the cryptocurrency of the ETC blockchain) is anomalously high compared to all the remaining parameters of the system. An encoder-decoder would interpret this quantity as noise and consequently would ignore such a value in the reconstruction. As a consequence, the di erence between the original and the reconstructed values would highlight the anomalous situation and consequently trigger the alert.
The model works on sequences of temporally-sorted events. Formally, we assume that the data is organized as a sequence X = f~x1; : : : ; ~xN g relative to a period of observation, where ~xt is a vector of features describing the t-th event in chronological order within X. An anomaly is an unexpected event in X, i.e. a vector ~xt signi cantly di erent from its neighbors ~xt w; : : : ; ~xt 1; ~xt+1; : : : ; ~xt+v (for a given window [t w; t + v] to be determined). Thus, the model should be capable of summarizing all the regularities in the data that characterize the sequence ~xt w; : : : ; ~xt+v.
Autoencoders [
The encoder , a neural network whose goal is to map an input ~x to a latent compact representation (~x) = ~z 2 RK . This mapping produces an embedding of the original input into a latent vector of size K.
A decoder , another neural networks that, given a K-dimensional vector ~z, aims at producing an output (~z) = ~y that is as close as possible to the original input.
In the above description, the components , , ~x and ~y are unspeci ed. In our framework,
we consider the input and output ~x; ~y of the autoencoder to be temporally marked events. As a
consequence and can be speci ed as recurrent networks (RNNs) [
~y (a) General autoencoder structure.
(b) Recurrent Autoencoder (RAE).
sequential data, since, iterating over the sequence, it saves (partial) memory of each event. In
our implementation we decided to instantiate the RNNs as Long short-term memory networks
(LSTM, see [
The resulting architecture is a sequence to sequence recurrent autoencoder (RAE) model [
n 1 X `(I; O) = n t=1 k~xt ~ytk2 where LSTM and mlp# represent the encoder, with internal state ~h(e) given the t-th event; t symmetrically, LSTM and mlp' represent the decoder, with inner state ~ht(d). Further, mlp# and mlp' represent multilayer networks parameterized by and , respectively.
Since the main purpose of RAE is to reconstruct the input from a compact representation, the model can be trained by considering a reconstruction loss:
Input subsequences are obtained from X through a sliding window mechanism. Each timestep within X is associated with a subsequence Wt = f~xt m+1; : : : ; : : : ; ~xtg, where m is the window size. Thus, RAE is trained on the set fWm; : : : ; WN g of subsequences that can obtained from X, as illustrated in Figure 2, by learning to reconstruct them in a way that minimizes the speci ed loss.
RAE will learn to model normal behavior, by mapping input data to compact representations that can capture all the relevant features characterizing it. An anomalous event will compromise this reconstruction capability. Thus, the distance between the input and the output can be exploited as a score to measure the outlierness of the analyzed sequence. Since each event can be included within several overlapping windows, its nal score is computed as the average of the outlierness scores of all the involved windows: where W~ i = dec(enc(Wi)) is the output of the RAE with input Wi. 4
The model proposed in the previous section represents a general approach to anomaly detection on sequential data. In this section, we apply our model to the ETC network to identify abnormal situations that can be symptoms of incoming attacks. As discussed in section 2, ETC has experienced two (known) successful attacks: the DAO attack (18 June 2016) and the 51% attack. The latter is scarcely documented and the reports state that it happened in a period within the interval 5-8 January 2019. Hence, our objective is to adopt the RAE model to check whether these events exhibit an anomaly score signi cantly higher than the other ones. 4.1
Our experiments have been conducted on a sample of ETC blockchain spanning over a period of four years (July 2015 - July 2019). The dataset is available on Kaggle2 and consists of seven tables, namely blocks, transactions, contracts, logs, token transfers, tokens and traces. These tables provide information about the usage of the network, the working status and the blocks.
A preliminary analysis of these tables highlighted that only two of them stored relevant information for our purposes, respectively blocks and transactions.
The raw data have been preprocessed as follows: (i) feature engineering, i.e. selection of the most relevant attribute, correlation analysis, ltering and aggregation; (ii) normalization; (iii) data view generation according to the sliding window mechanism.
2https://www.kaggle.com/bigquery/crypto-ethereum-classic
The rst two steps result in the following subset of relevant features, computed on a daily basis:
(i) block size average, the average size (in bytes) of a block; (ii) provided gas average, referring to the average provided gas needed to perform the transaction; (iii) block difficulty average, the average di culty (i.e. the e ort) necessary to validate a block; (iv) transaction average per block, the average number of transactions contained in a block; (v) gas used sum, the total amount of employed gas; (vi) transactions number, the total number of transactions in all blocks.
The normalization steps is meant to re-scale values to avoid instability e ects with the neural
modeling, as well as to regularize values by removing seasonal, cyclic and trending uctuations
in the data [
We partitioned the whole dataset in four periods as shown in Figure 3. Data prior to 7 August 2015 were ignored: they correspond to the rst week of the chain genesis and contain several missing values. The time series were nally produced for each block, to feed the RAE according to the learning/evaluation strategy outlined as follows. In practice, we consider two di erent training set, namely D1 (relative to the interval August 2015 - April 2016) and D2 (covering the interval July 2016 - November 2018). T1 and T2 (the periods within a range of about two months from the attacks, exact dates are listed in Figure 3) are used as test sets, for evaluating the outlierness scores. 4.2
We performed three di erent tests by training two instances of RAE. The rst two exploited D1 as training set and scored all the events within T1 and T2, respectively. In the third experiment RAE is trained on D1 [ D2 and scores events in T2.
Figures 4a and 4b plot the outlierness scores computed by RAE on both T1 and T2. We
can see that the DAO attack is perfectly detected by RAE, as shown in Figure 4a. The network
provides no warning until day 295 where we can see a small peak. Instead, the outlierness score
exhibits a di erence of two orders of magnitude on day 316, corresponding to the DAO attack.
The result is consistent with the ndings of [
Within Figure 4b, the same model is used to score events in T2. We can see that the
outlierness score changes its pattern as the attack period approaches. However, the peak is
translated of a few days. The situation represented here needs some interpretation. The
network highlights a strong anomaly on day 1255, a few days after the actual attack. A close
inspection of the events occurring during the 51% attack witnesses that a signi cant number
of companies, working on ETC, detected the attack and decided to freeze their activities3 4 5.
Then, on the basis of our analysis, block size and related features seems not fully su cient to
early detect some types of attacks occurring on blockchain, but we gure out that integrating
data from other sources (e.g. the server operating system and the application [
(h) transaction average per block norm 2 1 0 1 8 s a 6 G 4 2 0 activities were recorded on the network, thus representing the actual moment where the anomaly was injected within the blockchain. This nding is consistent with the ndings of the third experiment, where RAE is trained both on D1 and D2: Figure 4c shows a similar pattern, thus demonstrating that the preprocessing steps make the RAE ndings consistent even on di erent time periods. The only false positive seems to be represented by the peak on day 1291, which appears in both models. Although there is no reporting of attacks in those periods, it would be interesting to provide a close inspection to the activities during that period and try to explain such anomaly based on those. 15 7 10 5 15 2 re100 o c S 75 0 5 5 2 0 246 252 258 264 270 276 28 T2imesta3mp 2 88 294 00 306 312 318 324 330 336 342 348 0 3 re15 o c S 0 1 5 0
1200 1206 1212 1218 1224 1230 1236Ti1m242e1s24t8a1m254p1260 1266 1272 1278 1284 1290 1296 1302 (c) Outlierness score on T2. RAE is trained on
D1 [ D2.
In this work, we de ne an encoder-decoder deep learning model to detect anomalies in the usage of blockchain systems. In this scenario, the main problem is the rarity of the events to identify (i.e. the attacks), therefore, an unsupervised solution is adopted to address this issue. Speci cally, the contribution of the paper is twofold: (i) we identi ed a relevant set of features computed on blockchain logs describing network status in determined time steps; (ii) a sequence-to-sequence neural network model is used to recognize anomalous changes in the blockchain network. Experiments on the complete historical logs of Ethereum Classic prove our model capability to e ectively detect security attacks.
As future works, we plan to study the usage of hybrid architectures based on the combination
of RNNs with Convolutional Neural Networks [
This work has been partially supported by MIUR - PON Research and Innovation 2014-2020 under project Secure Open Nets.