BGBrunno Figueiroa Goldstein

Brunno Figueiroa Goldstein

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Brunno Figueiroa Goldstein has a total of 24 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Preventing DNN Model IP Theft via Hardware Obfuscation

    by Brunno F. Goldstein, Vinay C. Patil, Victor C. Ferreira, Alexandre S. Nery, Felipe M. G. Franca, Sandip Kundu

    Training accurate deep learning (DL) models require large amounts of training data, significant work in labeling the data, considerable computing resources, and substantial domain expertise. In short, they are expensive to develop. Hence, protecting these models, which are valuable storehouses of intellectual properties (IP), against model stealing/cloning attacks is of paramount importance. Today's mobile processors feature Neural Processing Units (NPUs) to accelerate the execution of DL models. DL models executing on NPUs are vulnerable to hyperparameter extraction via side-channel attacks and model parameter theft via bus monitoring attacks. This paper presents a novel solution to defend against DL IP theft in NPUs during model distribution and deployment/execution via lightweight, keyed model obfuscation scheme. Unauthorized use of such models results in inaccurate classification. In addition, we present an ideal end-to-end deep learning trusted system composed of: 1) model distribution via hardware root-of-trust and public-key cryptography infrastructure (PKI) and 2) model execution via low-latency memory encryption. We demonstrate that our proposed obfuscation solution achieves IP protection objectives without requiring specialized training or sacrificing the model's accuracy. In addition, the proposed obfuscation mechanism preserves the output class distribution while degrading the model's accuracy for unauthorized parties, covering any evidence of a hacked model.


  • Reliability Evaluation of Compressed Deep Learning Models

    by Brunno F. Goldstein, Sudarshan Srinivasan, Dipankar Das, Kunal Banerjee, Leandro Santiago, Victor C. Ferreira, Alexandre S. Nery, Sandip Kundu, Felipe M. G. Franca

    Neural networks are becoming deeper and more complex, making it harder to store and process such applications on systems with limited resources. Model pruning and data quantization are two effective ways to simplify the necessary hardware by compressing the network with relevant-only nodes and reducing the required data precision. Such optimizations, however, might come at a cost of reliability since critical nodes are now more exposed to faults and the network is more sensitive to small changes. In this work, we present an extensive empirical investigation of transient faults on compressed deep convolutional neural networks (CNNs). We evaluate the impact of a single bit flip over three CNN models with different sparsity configurations and integer-only quantizations. We show that pruning can increase the resilience of the system by 9x when compared to the dense model. Quantization can outperform the 32-bit floating-point baseline by adding 27.4x more resilience to the overall network and up to 108.7x when combined with pruning. This makes model compression an effective way to provide resilience to deep learning workloads during inference, mitigating the need for explicit error correction hardware.


  • Hardware-Accelerated Similarity Search with Multi-Index Hashing

    by Leandro Santiago de Araujo, Victor da Cruz Ferreira, Brunno Figueiroa Goldstein, Alexandre Solon Nery, Leandro Augusto Justen Marzulo, Sandip Kundu, Felipe Maia Galvao Franca

    Similarity search plays a major role in a large variety of database applications where high-dimensional data, usually derived from images and videos, are mapped onto indexed structures. Its essence relies on a k-nearest-neighbors (kNN) algorithm to find similar contents given a specific input data. An efficient way to store and speedup such searches is to apply kNN in Hamming space by encoding the data as binary and splitting it into multiple hash tables. This paper proposes an efficient Hardware-Accelerated Similarity Search co-processor architecture using Multi-Index Hashing as storage structure. The accelerator is specified in Verilog hardware description language and implemented in a Xilinx low-cost Zynq FPGA. Performance, circuit-area and power consumption results are presented, with the accelerator being up to 13x and 18x faster than the corresponding C/C++ code on the ARM host processor when running the SIFT and GIST datasets, respectively, while also requiring less power.

  • Teaching High Performance Computing through Parallel Programming Marathons

    by Leandro Marzulo, Calebe Bianchini, Leandro Santiago, Victor Ferreira, Brunno Goldstein, Felipe Franca

    Parallel and distributed programming is essential for exploiting the processing power of modern computing platforms. However, during the first years of a Computer Science course, students usually learn problem solving techniques, data structures and programming paradigms that are inherently sequential, hindering the transition to parallel architectures. Parallel Programming Marathons organized in Brazil are similar to other Programming Competitions around the world and have been used for teaching and stimulating undergraduate and graduate students into learning to "think in parallel" and to develop applications for different parallel architectures, including multicores, clusters and accelerators. This paper presents the structure of this Parallel Programming Marathon and an overview of how it supports regional and national contests. Also, this work presents use cases on Parallel and Distributed Computing course from two different Brazilian universities that use a challenge based learning approach and employ marathon problems as course assignments. This approach contributed to increase students' interest towards High Performance Computing.


  • A Feasible FPGA Weightless Neural Accelerator

    by Victor C. Ferreira, Alexandre S. Nery, Leandro A. J. Marzulo, Leandro Santiago, Diego Souza, Brunno F. Goldstein, Felipe M. G. Franca, Vladimir Alves

    AI applications have recently driven the computer architecture industry towards novel and more efficient dedicated hardware accelerators and tools. Weightless Neural Networks (WNNs) is a class of Artificial Neural Networks (ANNs) often applied to pattern recognition problems. It uses a set of Random Access Memories (RAMs) as the main mechanism for training and classifying information regarding a given input pattern. Due to its memory-based architecture, it can be easily mapped onto hardware and also greatly accelerated by means of a dedicated Register Transfer-Level (RTL) architecture designed to enable multiple memory accesses in parallel. On the other hand, a straightforward WNN hardware implementation requires too much memory resources for both ASIC and FPGA variants. This work aims at designing and evaluating a Weightless Neural accelerator designed in High-Level Synthesis (HLS). Our WNN accelerator implements Hash Tables, instead of regular RAMs, to substantially reduce its memory requirements, so that it can be implemented in a fairly small-sized Xilinx FPGA. Performance, circuit-area and power consumption results show that our accelerator can efficiently learn and classify the MNIST dataset in about 8 times faster than the system's embedded ARM processor.


  • Task Scheduling in Sucuri Dataflow Library

    by Rafael J.N. Silva, Brunno Goldstein, Leandro Santiago, Alexandre C. Sena, Leandro A.J. Marzulo, Tiago A.O. Alves, Felipe M.G. Franca

    Sucuri is a minimalistic Python library that provides dataflow programming through a reasonably simple syntax. It allows transparent execution on computer clusters and natural exploitation of parallelism. In Sucuri, programmers instantiate a dataflow graph, where each node is assigned to a function and edges represent data dependencies between nodes. The original implementation of Sucuri adopts a centralized scheduler, which incurs high communication overheads, specially in clusters with a large number of machines. In this paper we modify Sucuri so that each machine in a cluster will have its own scheduler. Before execution, the dataflow graph is partitioned, so that nodes can be distributed among the machines of the cluster. In runtime, idle workers will grab tasks from a ready queue in their local scheduler. Experimental results confirm that the solution can reduce communication overheads, improving performance in larger clusters.


  • Exploiting Parallelism in Linear Algebra Kernels through Dataflow Execution

    by Brunno F. Goldstein, Felipe M.G. Franca, Leandro A.J. Marzulo, Tiago A.O. Alves

    Linear Algebra Kernels have an important role in many petroleum reservoir simulators, extensively used by the industry. The growth in problem size, specially in pre-salt exploration, has caused an increase in execution time of those kernels, thus requiring parallel programming to improve performance and make the simulation viable. On the other hand, exploiting parallelism in systems with an ever increasing number of cores may be an arduous task, as the programmer has to manage threads and care about synchronization issues. Current work on parallel programming models show that Dataflow Execution exploits parallelism in a natural way, allowing the programmer to focus solely on describing dependencies between portions of code. This work consists in implementing parallel Linear Algebra Kernels using the Dataflow model. The Trebuchet Dataflow Virtual Machine and the Sucuri Dataflow Library were used to evaluate the solutions with real inputs from reservoir simulators. Results have been compared with OpenMP and Intel Math Kernel Library and show that coarser-grained tasks are needed to hide the overheads of dataflow runtime environments. Therefore, level 2 and 3 linear algebra operations, such as Vector-Matrix and Matrix-Matrix products, presented the most promising results.


  • Stack-Tagged Dataflow

    by Leandro Santiago, Leandro A.J. Marzulo, Brunno F. Goldstein, Tiago A.O. Alves, Felipe M.G. Franca

    Dynamic Dataflow allows simultaneous execution of instructions in different iterations of a loop, boosting parallelism exploitation. In this model, operands are tagged with their associated instance number, incremented as they go through the loop. Instruction execution is triggered when all input operand with the same tag become available. However, this traditional tagging mechanism often requires the generation of several control instructions to manipulate tags and guarantee the correct matching. To address this problem, this work presents Stack-Tagged Dataflow, a tagging mechanism that uses stacks of tags to reduce control overheads in dataflow, while isolating context from different nested loops. Push instructions at the beginning of loops will create a new context, while Pop instructions, at the end of the loop, will restore the original context tags. Experimental results show that Stack-Tagged Dataflow is a viable solution, suggesting that a hybrid compiling approach can be used to enable Stack-Tags only when needed.


  • A Minimalistic Dataflow Programming Library for Python

    by Tiago A.O. Alves, Brunno F. Goldstein, Felipe M.G. Franca, Leandro A.J. Marzulo

    Current work on parallel programming models are trending towards the dataflow paradigm. Previous works on that topic have shown that dataflow programming is indeed a natural way to exploit parallelism in programs. However, there is still a gap in terms of ease of programming between high level languages adopted by the scientific community and the languages and tools available for dataflow programming. In this paper we present Sucuri: a minimalistic Python library that provides dataflow programming with reasonably simple syntax. To parallelize applications using our library, the programmer needs only to identify functions of his code that are good candidates for parallelization and instantiate a dataflow graph where each node is associated with one of such functions, and the edges between nodes describe data dependencies between functions. We then proceed to implement two benchmarks that represent important parallel programming patterns using our library and execute them on a cluster of multicores. Experimental results are promising, proving that our library can be an interesting first option for parallelization.